axolotl/search.json

[
  {
    "objectID": "FAQS.html",
    "href": "FAQS.html",
    "title": "FAQs",
    "section": "",
    "text": "FAQs\n\nCan you train StableLM with this? Yes, but only with a single GPU atm. Multi GPU support is coming soon! Just waiting on this PR\nWill this work with Deepspeed? That’s still a WIP, but setting export ACCELERATE_USE_DEEPSPEED=true should work in some cases\nError invalid argument at line 359 in file /workspace/bitsandbytes/csrc/pythonInterface.c\n/arrow/cpp/src/arrow/filesystem/s3fs.cc:2598:  arrow::fs::FinalizeS3 was not called even though S3 was initialized.\nThis could lead to a segmentation fault at exit. Try reinstalling bitsandbytes and transformers from source."
  },
  {
    "objectID": "src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html",
    "href": "src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html",
    "title": "Axolotl",
    "section": "",
    "text": "Acknowledgements\nPortions of this Cut Cross Entropy Software may utilize the following copyrighted\nmaterial, the use of which is hereby acknowledged.\n\nPyTorch\nFrom PyTorch:\n\nCopyright (c) 2016-     Facebook, Inc            (Adam Paszke)\nCopyright (c) 2014-     Facebook, Inc            (Soumith Chintala)\nCopyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)\nCopyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)\nCopyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)\nCopyright (c) 2011-2013 NYU                      (Clement Farabet)\nCopyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)\nCopyright (c) 2006      Idiap Research Institute (Samy Bengio)\nCopyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)\n\nFrom Caffe2:\n\nCopyright (c) 2016-present, Facebook Inc. All rights reserved.\n\nAll contributions by Facebook:\nCopyright (c) 2016 Facebook Inc.\n\nAll contributions by Google:\nCopyright (c) 2015 Google Inc.\nAll rights reserved.\n\nAll contributions by Yangqing Jia:\nCopyright (c) 2015 Yangqing Jia\nAll rights reserved.\n\nAll contributions by Kakao Brain:\nCopyright 2019-2020 Kakao Brain\n\nAll contributions by Cruise LLC:\nCopyright (c) 2022 Cruise LLC.\nAll rights reserved.\n\nAll contributions by Arm:\nCopyright (c) 2021, 2023-2024 Arm Limited and/or its affiliates\n\nAll contributions from Caffe:\nCopyright(c) 2013, 2014, 2015, the respective contributors\nAll rights reserved.\n\nAll other contributions:\nCopyright(c) 2015, 2016 the respective contributors\nAll rights reserved.\n\nCaffe2 uses a copyright model similar to Caffe: each contributor holds\ncopyright over their contributions to Caffe2. The project versioning records\nall such contribution and copyright details. If a contributor wants to further\nmark their specific copyright on a particular contribution, they should\nindicate their copyright solely in the commit message of the change when it is\ncommitted.\n\nAll rights reserved.\n\nRedistribution and use in source and binary forms, with or without\nmodification, are permitted provided that the following conditions are met:\n\n1. Redistributions of source code must retain the above copyright\nnotice, this list of conditions and the following disclaimer.\n\n2. Redistributions in binary form must reproduce the above copyright\nnotice, this list of conditions and the following disclaimer in the\ndocumentation and/or other materials provided with the distribution.\n\n3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America\nand IDIAP Research Institute nor the names of its contributors may be\nused to endorse or promote products derived from this software without\nspecific prior written permission.\n\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\"\nAND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE\nIMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE\nARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE\nLIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR\nCONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF\nSUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS\nINTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN\nCONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)\nARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE\nPOSSIBILITY OF SUCH DAMAGE.\nTriton\n/*\n* Copyright 2018-2020 Philippe Tillet\n* Copyright 2020-2022 OpenAI\n*\n* Permission is hereby granted, free of charge, to any person obtaining\n* a copy of this software and associated documentation files\n* (the \"Software\"), to deal in the Software without restriction,\n* including without limitation the rights to use, copy, modify, merge,\n* publish, distribute, sublicense, and/or sell copies of the Software,\n* and to permit persons to whom the Software is furnished to do so,\n* subject to the following conditions:\n*\n* The above copyright notice and this permission notice shall be\n* included in all copies or substantial portions of the Software.\n*\n* THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND,\n* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF\n* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.\n* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY\n* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,\n* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE\n* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\n*/\nTransformers\nCopyright 2018- The Hugging Face team. All rights reserved.\n\n                                Apache License\n                        Version 2.0, January 2004\n                        http://www.apache.org/licenses/\n\nTERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION\n\n1. Definitions.\n\n    \"License\" shall mean the terms and conditions for use, reproduction,\n    and distribution as defined by Sections 1 through 9 of this document.\n\n    \"Licensor\" shall mean the copyright owner or entity authorized by\n    the copyright owner that is granting the License.\n\n    \"Legal Entity\" shall mean the union of the acting entity and all\n    other entities that control, are controlled by, or are under common\n    control with that entity. For the purposes of this definition,\n    \"control\" means (i) the power, direct or indirect, to cause the\n    direction or management of such entity, whether by contract or\n    otherwise, or (ii) ownership of fifty percent (50%) or more of the\n    outstanding shares, or (iii) beneficial ownership of such entity.\n\n    \"You\" (or \"Your\") shall mean an individual or Legal Entity\n    exercising permissions granted by this License.\n\n    \"Source\" form shall mean the preferred form for making modifications,\n    including but not limited to software source code, documentation\n    source, and configuration files.\n\n    \"Object\" form shall mean any form resulting from mechanical\n    transformation or translation of a Source form, including but\n    not limited to compiled object code, generated documentation,\n    and conversions to other media types.\n\n    \"Work\" shall mean the work of authorship, whether in Source or\n    Object form, made available under the License, as indicated by a\n    copyright notice that is included in or attached to the work\n    (an example is provided in the Appendix below).\n\n    \"Derivative Works\" shall mean any work, whether in Source or Object\n    form, that is based on (or derived from) the Work and for which the\n    editorial revisions, annotations, elaborations, or other modifications\n    represent, as a whole, an original work of authorship. For the purposes\n    of this License, Derivative Works shall not include works that remain\n    separable from, or merely link (or bind by name) to the interfaces of,\n    the Work and Derivative Works thereof.\n\n    \"Contribution\" shall mean any work of authorship, including\n    the original version of the Work and any modifications or additions\n    to that Work or Derivative Works thereof, that is intentionally\n    submitted to Licensor for inclusion in the Work by the copyright owner\n    or by an individual or Legal Entity authorized to submit on behalf of\n    the copyright owner. For the purposes of this definition, \"submitted\"\n    means any form of electronic, verbal, or written communication sent\n    to the Licensor or its representatives, including but not limited to\n    communication on electronic mailing lists, source code control systems,\n    and issue tracking systems that are managed by, or on behalf of, the\n    Licensor for the purpose of discussing and improving the Work, but\n    excluding communication that is conspicuously marked or otherwise\n    designated in writing by the copyright owner as \"Not a Contribution.\"\n\n    \"Contributor\" shall mean Licensor and any individual or Legal Entity\n    on behalf of whom a Contribution has been received by Licensor and\n    subsequently incorporated within the Work.\n\n2. Grant of Copyright License. Subject to the terms and conditions of\n    this License, each Contributor hereby grants to You a perpetual,\n    worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n    copyright license to reproduce, prepare Derivative Works of,\n    publicly display, publicly perform, sublicense, and distribute the\n    Work and such Derivative Works in Source or Object form.\n\n3. Grant of Patent License. Subject to the terms and conditions of\n    this License, each Contributor hereby grants to You a perpetual,\n    worldwide, non-exclusive, no-charge, royalty-free, irrevocable\n    (except as stated in this section) patent license to make, have made,\n    use, offer to sell, sell, import, and otherwise transfer the Work,\n    where such license applies only to those patent claims licensable\n    by such Contributor that are necessarily infringed by their\n    Contribution(s) alone or by combination of their Contribution(s)\n    with the Work to which such Contribution(s) was submitted. If You\n    institute patent litigation against any entity (including a\n    cross-claim or counterclaim in a lawsuit) alleging that the Work\n    or a Contribution incorporated within the Work constitutes direct\n    or contributory patent infringement, then any patent licenses\n    granted to You under this License for that Work shall terminate\n    as of the date such litigation is filed.\n\n4. Redistribution. You may reproduce and distribute copies of the\n    Work or Derivative Works thereof in any medium, with or without\n    modifications, and in Source or Object form, provided that You\n    meet the following conditions:\n\n    (a) You must give any other recipients of the Work or\n        Derivative Works a copy of this License; and\n\n    (b) You must cause any modified files to carry prominent notices\n        stating that You changed the files; and\n\n    (c) You must retain, in the Source form of any Derivative Works\n        that You distribute, all copyright, patent, trademark, and\n        attribution notices from the Source form of the Work,\n        excluding those notices that do not pertain to any part of\n        the Derivative Works; and\n\n    (d) If the Work includes a \"NOTICE\" text file as part of its\n        distribution, then any Derivative Works that You distribute must\n        include a readable copy of the attribution notices contained\n        within such NOTICE file, excluding those notices that do not\n        pertain to any part of the Derivative Works, in at least one\n        of the following places: within a NOTICE text file distributed\n        as part of the Derivative Works; within the Source form or\n        documentation, if provided along with the Derivative Works; or,\n        within a display generated by the Derivative Works, if and\n        wherever such third-party notices normally appear. The contents\n        of the NOTICE file are for informational purposes only and\n        do not modify the License. You may add Your own attribution\n        notices within Derivative Works that You distribute, alongside\n        or as an addendum to the NOTICE text from the Work, provided\n        that such additional attribution notices cannot be construed\n        as modifying the License.\n\n    You may add Your own copyright statement to Your modifications and\n    may provide additional or different license terms and conditions\n    for use, reproduction, or distribution of Your modifications, or\n    for any such Derivative Works as a whole, provided Your use,\n    reproduction, and distribution of the Work otherwise complies with\n    the conditions stated in this License.\n\n5. Submission of Contributions. Unless You explicitly state otherwise,\n    any Contribution intentionally submitted for inclusion in the Work\n    by You to the Licensor shall be under the terms and conditions of\n    this License, without any additional terms or conditions.\n    Notwithstanding the above, nothing herein shall supersede or modify\n    the terms of any separate license agreement you may have executed\n    with Licensor regarding such Contributions.\n\n6. Trademarks. This License does not grant permission to use the trade\n    names, trademarks, service marks, or product names of the Licensor,\n    except as required for reasonable and customary use in describing the\n    origin of the Work and reproducing the content of the NOTICE file.\n\n7. Disclaimer of Warranty. Unless required by applicable law or\n    agreed to in writing, Licensor provides the Work (and each\n    Contributor provides its Contributions) on an \"AS IS\" BASIS,\n    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or\n    implied, including, without limitation, any warranties or conditions\n    of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A\n    PARTICULAR PURPOSE. You are solely responsible for determining the\n    appropriateness of using or redistributing the Work and assume any\n    risks associated with Your exercise of permissions under this License.\n\n8. Limitation of Liability. In no event and under no legal theory,\n    whether in tort (including negligence), contract, or otherwise,\n    unless required by applicable law (such as deliberate and grossly\n    negligent acts) or agreed to in writing, shall any Contributor be\n    liable to You for damages, including any direct, indirect, special,\n    incidental, or consequential damages of any character arising as a\n    result of this License or out of the use or inability to use the\n    Work (including but not limited to damages for loss of goodwill,\n    work stoppage, computer failure or malfunction, or any and all\n    other commercial damages or losses), even if such Contributor\n    has been advised of the possibility of such damages.\n\n9. Accepting Warranty or Additional Liability. While redistributing\n    the Work or Derivative Works thereof, You may choose to offer,\n    and charge a fee for, acceptance of support, warranty, indemnity,\n    or other liability obligations and/or rights consistent with this\n    License. However, in accepting such obligations, You may act only\n    on Your own behalf and on Your sole responsibility, not on behalf\n    of any other Contributor, and only if You agree to indemnify,\n    defend, and hold each Contributor harmless for any liability\n    incurred by, or claims asserted against, such Contributor by reason\n    of your accepting any such warranty or additional liability.\n\nEND OF TERMS AND CONDITIONS\n\nAPPENDIX: How to apply the Apache License to your work.\n\n    To apply the Apache License to your work, attach the following\n    boilerplate notice, with the fields enclosed by brackets \"[]\"\n    replaced with your own identifying information. (Don't include\n    the brackets!)  The text should be enclosed in the appropriate\n    comment syntax for the file format. We also recommend that a\n    file or class name and description of purpose be included on the\n    same \"printed page\" as the copyright notice for easier\n    identification within third-party archives.\n\nCopyright [yyyy] [name of copyright owner]\n\nLicensed under the Apache License, Version 2.0 (the \"License\");\nyou may not use this file except in compliance with the License.\nYou may obtain a copy of the License at\n\n    http://www.apache.org/licenses/LICENSE-2.0\n\nUnless required by applicable law or agreed to in writing, software\ndistributed under the License is distributed on an \"AS IS\" BASIS,\nWITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\nSee the License for the specific language governing permissions and\nlimitations under the License."
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html",
    "href": "examples/colab-notebooks/colab-axolotl-example.html",
    "title": "Setting up",
    "section": "",
    "text": "import torch\n# Check so there is a gpu available, a T4(free tier) is enough to run this notebook\nassert (torch.cuda.is_available()==True)\n!pip install --no-build-isolation axolotl[deepspeed]"
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html#hugging-face-login-optional",
    "href": "examples/colab-notebooks/colab-axolotl-example.html#hugging-face-login-optional",
    "title": "Setting up",
    "section": "Hugging Face login (optional)",
    "text": "Hugging Face login (optional)\n\nfrom huggingface_hub import notebook_login\nnotebook_login()"
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html#example-configuration",
    "href": "examples/colab-notebooks/colab-axolotl-example.html#example-configuration",
    "title": "Setting up",
    "section": "Example configuration",
    "text": "Example configuration\n\nimport yaml\n\nyaml_string = \"\"\"\nbase_model: NousResearch/Meta-Llama-3.1-8B\n\nload_in_8bit: false\nload_in_4bit: true\nstrict: false\n\ndatasets:\n  - path: tatsu-lab/alpaca\n    type: alpaca\ndataset_prepared_path: last_run_prepared\nval_set_size: 0.05\noutput_dir: ./outputs/lora-out\n\nsequence_len: 2048\nsample_packing: true\neval_sample_packing: true\npad_to_sequence_len: true\n\nadapter: qlora\nlora_model_dir:\nlora_r: 32\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_linear: true\nlora_fan_in_fan_out:\nlora_modules_to_save:\n  - embed_tokens\n  - lm_head\n\nwandb_project:\nwandb_entity:\nwandb_watch:\nwandb_name:\nwandb_log_model:\n\ngradient_accumulation_steps: 2\nmicro_batch_size: 1\nnum_epochs: 1\noptimizer: paged_adamw_8bit\nlr_scheduler: cosine\nlearning_rate: 2e-5\n\ntrain_on_inputs: false\ngroup_by_length: false\nbf16: auto\nfp16:\ntf32: false\n\ngradient_checkpointing: true\nearly_stopping_patience:\nresume_from_checkpoint:\nlogging_steps: 1\nxformers_attention:\nflash_attention: false\nsdp_attention: true\n\nwarmup_steps: 1\nmax_steps: 25\nevals_per_epoch: 1\neval_table_size:\nsaves_per_epoch: 1\ndebug:\ndeepspeed:\nweight_decay: 0.0\nfsdp:\nfsdp_config:\nspecial_tokens:\n  pad_token: <|end_of_text|>\n\"\"\"\n\n\n# Convert the YAML string to a Python dictionary\nyaml_dict = yaml.safe_load(yaml_string)\n\n# Specify your file path\nfile_path = 'test_axolotl.yaml'\n\n# Write the YAML file\nwith open(file_path, 'w') as file:\n    yaml.dump(yaml_dict, file)\n\nAbove we have a configuration file with base LLM model and datasets specified, among many other things. Axolotl can automatically detect whether the specified datasets are on HuggingFace repo or local machine.\nThe Axolotl configuration options encompass model and dataset selection, data pre-processing, and training. Let’s go through them line by line:\n\n“base model”: String value, specifies the underlying pre-trained LLM that will be used for finetuning\n\nNext we have options for model weights quantization. Quantization allows for reduction in occupied memory on GPUs.\n\n“load_in_8bit”: Boolean value, whether to quantize the model weights into 8-bit integer.\n“load_in_4bit”: Boolean value, whether to quantize the model weights into 4-bit integer.\n“strict”: Boolean value. If false, it allows for overriding established configuration options in the yaml file when executing in command-line interface.\n“datasets”: a list of dicts that contain path and type of data sets as well as other optional configurations where datasets are concerned. Supports multiple datasets.\n“val_set_size”: Either a float value less than one or an integer less than the total size of dataset. Sets the size of validation set from the whole dataset. If float, sets the proportion of the dataset assigned for validation. If integer, sets the direct size of validation set.\n“output_dir”: String value. Path of trained model.\n\nFor data preprocessing:\n\n“sequence_len”: Integer. Specifies the maximum sequence length of the input. Typically 2048 or less.\n“pad_to_sequence_len”: Boolean. Padding input to maximum sequence length.\n“sample_packing”: Boolean. Specifies whether to use multi-packing with block diagonal attention.\n“special_tokens”: Python dict, optional. Allows users to specify the additional special tokens to be ignored by the tokenizer.\n\nFor LoRA configuration and its hyperparamters:\n\n“adapter”: String. Either “lora” or “qlora”, depending on user’s choice.\n“lora_model_dir”: String, Optional. Path to directory that contains LoRA model, if there is already a trained LoRA model the user would like to use.\n“lora_r”: Integer. Refers to the rank of LoRA decomposition matrices. Higher value will reduce LoRA efficiency. Recommended to be set to 8.\n“lora_alpha”: Integer. Scale the weight matrices by \\(\\frac{\\text{lora_alpha}}{\\text{lora_r}}\\)Recommended to be fixed at 16.\n“lora_dropout”: Float that is 1 or less. The dropout probability of a lora layer.\n“lora_target_linear”: Boolean. If true, lora will target all linear modules in the transformers architecture.\n“lora_modules_to_save”: If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n\nSee LoRA for detailed explanation of LoRA implementation.\nFor the training configurations:\n\n“gradient_accumulation_steps”: Integer. The number of steps over which to accumulate gradient for batch training. E.g. if 2, backprop is performed every two steps.\n“micro_batch_size”: Integer. Batch size per gpu / gradient_accumulation_steps\n“num_epochs”: Integer. Number of epochs. One epoch is when training has looped over every batch in the whole data set once.\n“optimizer”: The optimizer to use for the training.\n“learning_rate”: The learning rate.\n“lr_scheduler”: The learning rate scheduler to use for adjusting learning rate during training.\n“train_on_inputs”: Boolean. Whether to ignore or include the user’s prompt from the training labels.\n“group_by_length”: Boolean. Whether to group similarly sized data to minimize padding.\n“bf16”: Either “auto”, “true”, or “false”. Whether to use CUDA bf16 floating point format. If set to “auto”, will automatically apply bf16 should the gpu supports it.\n“fp16”: Optional. Specifies whether to use CUDA fp16. Automatically set to true if “bf16” is set to true. Otherwise false.\n“tf32”: Boolean. Whether to use CUDA tf32. Will override bf16.\n“gradient_checkpointing”: Boolean. Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\n“gradient_checkpointing_kwargs”: Python Dict. Fed into the trainer.\n“logging_steps”: Integer. Log training information over every specified number of steps.\n“flash_attention”: Boolean. Whether to use the flash attention mechanism.\n“sdp_attention”: Boolean. Whether to use the Scaled Dot Product attention mechanism (the attention mechanism in the original implementation of transformers.)\n“warmup_steps”: Integer. The number of pre-training steps where a very low learning rate is used.\n“evals_per_epoch”: Integer. Number of evaluations to be performed within one training epoch.\n“saves_per_epoch”: Integer. Number of times the model is saved in one training epoch.\n“weight_decay”: Positive Float. Sets the “strength” of weight decay (i.e. setting the coefficient of L2 regularization)\n\nThe above is but a snippet aiming to get users familiarized with the types of streamlined configuration options axolotl provides. For a full list of configuration options, see here\nTrain the model\n\n!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml\n\nPredict with trained model\n\n!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n    --lora_model_dir=\"./outputs/lora-out\" --gradio"
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html#deeper-dive",
    "href": "examples/colab-notebooks/colab-axolotl-example.html#deeper-dive",
    "title": "Setting up",
    "section": "Deeper Dive",
    "text": "Deeper Dive\nIt is also helpful to gain some familiarity over some of the core inner workings of axolotl"
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html#configuration-normalization",
    "href": "examples/colab-notebooks/colab-axolotl-example.html#configuration-normalization",
    "title": "Setting up",
    "section": "Configuration Normalization",
    "text": "Configuration Normalization\nAxolotl uses a custom Dict class, called DictDefault\nto store configurations specified in the yaml configuration file (into a Python variable named cfg). The definition for this custom Dict can be found in the utils/dict.py\nDictDefault is amended such that calling a missing key from it will result in a None return type. This is important because if some configuration options aren’t specified by the user, the None type allows Axolotl to perform boolean operations to determine the default settings for missing configurations. For more examples on how this is done, check out utils/config/init.py"
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html#loading-models-tokenizers-and-trainer",
    "href": "examples/colab-notebooks/colab-axolotl-example.html#loading-models-tokenizers-and-trainer",
    "title": "Setting up",
    "section": "Loading Models, Tokenizers, and Trainer",
    "text": "Loading Models, Tokenizers, and Trainer\nIf we inspect cli.train.py, we will find that most of the heavy lifting were done by the function train() which is itself imported from src/axolotl/train.py.\ntrain() takes care of loading the appropriate tokenizer and pre-trained model through load_model() and load_tokenizer() from src/axolotl/utils/models.py respectively.\nload_tokenizer() loads in the appropriate tokenizer given the desired model, as well as chat templates.\nModelLoader class follows after tokenizer has been selected. It will automatically discern the base model type, load in the desired model, as well as applying model-appropriate attention mechanism modifications (e.g. flash attention). Depending on which base model the user chooses in the configuration, ModelLoader will utilize the corresponding “attention hijacking” script. For example, if the user specified the base model to be NousResearch/Meta-Llama-3.1-8B, which is of llama type, and set flash_attn to True, ModelLoader will load in llama_attn_hijack_flash.py. For a list of supported attention hijacking, please refer to the directory /src/axolotl/monkeypatch/\nAnother important operation encompassed in train() is setting up the training that takes into account of user-specified traning configurations (e.g. num_epochs, optimizer) through the use of setup_trainer() from /src/axolotl/utils/trainer.py, which in turn relies on modules from /src/axolotl/core/trainer_builder.py.\ntrainer_builder.py provides a list of trainer object options bespoke for the task type (Causal or Reinforcement learning (‘dpo’, ‘ipo’, ‘kto’) )"
  },
  {
    "objectID": "examples/colab-notebooks/colab-axolotl-example.html#monkey-patch",
    "href": "examples/colab-notebooks/colab-axolotl-example.html#monkey-patch",
    "title": "Setting up",
    "section": "Monkey patch",
    "text": "Monkey patch\nThe Monkey patch directory is where model architecture/optimization patching scripts are stored (these are modifications that are not implemented in the official releases, hence the name monkey patch). It includes attention jacking, ReLoRA, and unsloth optimization."
  },
  {
    "objectID": "docs/lr_groups.html",
    "href": "docs/lr_groups.html",
    "title": "Learning Rate Groups",
    "section": "",
    "text": "Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of\nmodules in a model.",
    "crumbs": [
      "How To Guides",
      "Learning Rate Groups"
    ]
  },
  {
    "objectID": "docs/lr_groups.html#background",
    "href": "docs/lr_groups.html#background",
    "title": "Learning Rate Groups",
    "section": "",
    "text": "Inspired by LoRA+, Axolotl allows practitioners to specify separate learning rates for each module or groups of\nmodules in a model.",
    "crumbs": [
      "How To Guides",
      "Learning Rate Groups"
    ]
  },
  {
    "objectID": "docs/lr_groups.html#example",
    "href": "docs/lr_groups.html#example",
    "title": "Learning Rate Groups",
    "section": "Example",
    "text": "Example\nlr_groups:\n  - name: o_proj\n    modules:\n      - self_attn.o_proj.weight\n    lr: 1e-6\n  - name: q_proj\n    modules:\n      - model.layers.2.self_attn.q_proj.weight\n    lr: 1e-5\n\nlearning_rate: 2e-5\nIn this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate\nof 1e-6 for all the self attention o_proj modules across all layers, and a learning are of 1e-5 to the 3rd layer’s\nself attention q_proj module.",
    "crumbs": [
      "How To Guides",
      "Learning Rate Groups"
    ]
  },
  {
    "objectID": "docs/lora_optims.html",
    "href": "docs/lora_optims.html",
    "title": "LoRA Optimizations",
    "section": "",
    "text": "Inspired by Unsloth, we’ve implemented two\noptimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU\n(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function\nTriton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was\nto leverage operator fusion and tensor re-use in order to improve speed and reduce\nmemory usage during the forward and backward passes of these calculations.\nWe currently support several common model architectures, including (but not limited to):",
    "crumbs": [
      "How To Guides",
      "LoRA Optimizations"
    ]
  },
  {
    "objectID": "docs/lora_optims.html#usage",
    "href": "docs/lora_optims.html#usage",
    "title": "LoRA Optimizations",
    "section": "Usage",
    "text": "Usage\nThese optimizations can be enabled in your Axolotl config YAML file. The\nlora_mlp_kernel option enables the optimized MLP path, while lora_qkv_kernel and\nlora_o_kernel enable the fused query-key-value projection and optimized output\nprojection, respectively.\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true\n\n\n\n\n\n\nNote\n\n\n\nCurrently, LoRA kernels are not supported for RLHF training, only SFT.",
    "crumbs": [
      "How To Guides",
      "LoRA Optimizations"
    ]
  },
  {
    "objectID": "docs/lora_optims.html#requirements",
    "href": "docs/lora_optims.html#requirements",
    "title": "LoRA Optimizations",
    "section": "Requirements",
    "text": "Requirements\n\nOne or more NVIDIA or AMD GPUs (in order to use the Triton kernels)\n\nNote: Set TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 to enable memory-efficient attention on AMD GPUs\n\nTargeted LoRA adapters cannot use Dropout\n\nThis may limit model expressivity / cause overfitting\n\nTargeted LoRA adapters cannot have bias terms\n\nThis may limit model expressivity\n\n\nModels with pre-existing LoRA adapters that use Dropout or have bias terms may need to\nbe re-finetuned without these features in order to be useful.",
    "crumbs": [
      "How To Guides",
      "LoRA Optimizations"
    ]
  },
  {
    "objectID": "docs/lora_optims.html#implementation-details",
    "href": "docs/lora_optims.html#implementation-details",
    "title": "LoRA Optimizations",
    "section": "Implementation details",
    "text": "Implementation details\n\nCustom autograd functions\nThe LoRA MLP autograd function optimizes the entire MLP computation path. It fuses the\nLoRA and base weight computations together and provides a single, efficient backward\npass for the entire MLP block.\nFor attention components, similar optimizations are provided through a function that\nhandles the query, key, and value projections, and a function that handles the output\nprojection. They are designed to work with the existing transformers attention\nimplementation via some monkey-patching logic.\n\n\nTriton kernels\nTwo activation functions (SwiGLU and GeGLU) are implemented with Triton kernels for\nimproved speed and memory performance. These kernels handle both the forward and\nbackward passes.\n\n\nIntegration\nThe custom autograd functions and Triton kernels are designed to work together. The\nautograd function manages the high-level computation flow and gradient tracking, while\ncalling the Triton kernels for the activation function computation. During the backward\npass, the kernel computes both the activation output and the required gradients, which\nthe autograd function then uses to compute the final gradients for the entire\ncomputation path.",
    "crumbs": [
      "How To Guides",
      "LoRA Optimizations"
    ]
  },
  {
    "objectID": "docs/lora_optims.html#future-work",
    "href": "docs/lora_optims.html#future-work",
    "title": "LoRA Optimizations",
    "section": "Future Work",
    "text": "Future Work\n\nSupport for additional model architectures\nSupport for the FSDP setting\nSupport for dropout and bias\nAdditional operator fusions",
    "crumbs": [
      "How To Guides",
      "LoRA Optimizations"
    ]
  },
  {
    "objectID": "docs/rlhf.html",
    "href": "docs/rlhf.html",
    "title": "RLHF (Beta)",
    "section": "",
    "text": "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human\nfeedback. Various methods include, but not limited to:\n\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)\nProximal Policy Optimization (PPO) (not yet supported in axolotl, if you’re interested in contributing, please reach out!)",
    "crumbs": [
      "How To Guides",
      "RLHF (Beta)"
    ]
  },
  {
    "objectID": "docs/rlhf.html#overview",
    "href": "docs/rlhf.html#overview",
    "title": "RLHF (Beta)",
    "section": "",
    "text": "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human\nfeedback. Various methods include, but not limited to:\n\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)\nProximal Policy Optimization (PPO) (not yet supported in axolotl, if you’re interested in contributing, please reach out!)",
    "crumbs": [
      "How To Guides",
      "RLHF (Beta)"
    ]
  },
  {
    "objectID": "docs/rlhf.html#rlhf-using-axolotl",
    "href": "docs/rlhf.html#rlhf-using-axolotl",
    "title": "RLHF (Beta)",
    "section": "RLHF using Axolotl",
    "text": "RLHF using Axolotl\n\n\n\n\n\n\nImportant\n\n\n\nThis is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.\n\n\nWe rely on the TRL library for implementations of various RL training methods, which we wrap around to expose in axolotl. Each method has their own supported ways of loading datasets and prompt formats.\n\n\n\n\n\n\nTip\n\n\n\nYou can find what each method supports by going into src/axolotl/prompt_strategies/{method} where {method} is one of our supported methods. The type: can be retrieved from {method}.{function_name}.\n\n\n\nDPO\nExample config:\nrl: dpo\ndatasets:\n  - path: Intel/orca_dpo_pairs\n    split: train\n    type: chatml.intel\n  - path: argilla/ultrafeedback-binarized-preferences\n    split: train\n    type: chatml\nDPO supports the following types with the following dataset format:\n\nchatml.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"chosen_response\": \"...\",\n    \"rejected_response\": \"...\"\n}\n\n\nchatml.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nchatml.icr\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nchatml.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"chosen_response\": \"...\",\n    \"rejected_response\": \"...\"\n}\n\n\nllama3.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.icr\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nllama3.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nzephyr.nectar\n{\n    \"prompt\": \"...\",\n    \"answers\": [\n        {\n            \"answer\": \"...\",\n            \"rank\": 1\n        },\n        {\n            \"answer\": \"...\",\n            \"rank\": 2\n        }\n        // ... more answers with ranks\n    ]\n}\n\n\nchat_template.default\nrl: dpo\ndatasets:\n  - path: ...\n    split: train\n    type: chat_template.default\n    field_messages: \"messages\"\n    field_chosen: \"chosen\"\n    field_rejected: \"rejected\"\n    message_property_mappings:\n      role: role\n      content: content\n    roles:\n      user: [\"user\"]\n      assistant: [\"assistant\"]\n      system: [\"system\"]\nSample input format:\n{\n    \"messages\": [\n        {\n            \"role\": \"system\",\n            \"content\": \"...\"\n        },\n        {\n            \"role\": \"user\",\n            \"content\": \"...\"\n        },\n        // ... more messages\n    ],\n    \"chosen\": {\n        \"role\": \"assistant\",\n        \"content\": \"...\"\n    },\n    \"rejected\": {\n        \"role\": \"assistant\",\n        \"content\": \"...\"\n    }\n}\n\n\nuser_defined.default\nFor custom behaviors,\nrl: dpo\ndatasets:\n  - path: ...\n    split: train\n    type: user_defined.default\n\n    field_prompt: \"prompt\"\n    field_system: \"system\"\n    field_chosen: \"chosen\"\n    field_rejected: \"rejected\"\n    prompt_format: \"{prompt}\"\n    chosen_format: \"{chosen}\"\n    rejected_format: \"{rejected}\"\nThe input format is a simple JSON input with customizable fields based on the above config.\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\n\nIPO\nAs IPO is just DPO with a different loss function, all supported dataset formats for DPO are also supported for IPO.\nrl: ipo\n\n\nORPO\nPaper: https://arxiv.org/abs/2403.07691\nrl: orpo\norpo_alpha: 0.1\nremove_unused_columns: false\n\nchat_template: chatml\ndatasets:\n  - path: argilla/ultrafeedback-binarized-preferences-cleaned\n    type: chat_template.argilla\nORPO supports the following types with the following dataset format:\n\nchat_template.argilla\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",  // if available, will be taken as user message for single-turn instead of from list below\n\n    // chosen/rejected should be same till last content and only even-number of alternating user/assistant turns\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ],\n    \"rejected\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\n\nKTO\nrl: kto\nrl_beta: 0.1  # default\nkto_desirable_weight: 1.0  # default\nkto_undesirable_weight: 1.0  # default\n\nremove_unused_columns: false\n\ndatasets:\n  - path: argilla/ultrafeedback-binarized-preferences-cleaned-kto\n    type: llama3.ultra\n    split: train\n\ngradient_checkpointing: true\ngradient_checkpointing_kwargs:\n  use_reentrant: true\nKTO supports the following types with the following dataset format:\n\nchatml.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.argilla_chat\n{\n    \"chosen\": [\n        {\"role\": \"user\", \"content\": \"...\"}\n    ],\n    \"completion\": [\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nchatml.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nchatml.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.argilla\n{\n    \"system\": \"...\", // optional\n    \"instruction\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.argilla_chat\n{\n    \"completion\": [\n        {\"role\": \"user\", \"content\": \"...\"},\n        {\"role\": \"assistant\", \"content\": \"...\"}\n    ]\n}\n\n\nllama3.intel\n{\n    \"system\": \"...\", // optional\n    \"question\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.prompt_pairs\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nllama3.ultra\n{\n    \"system\": \"...\", // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\"\n}\n\n\nuser_defined.default\nFor custom behaviors,\nrl: kto\ndatasets:\n  - path: ...\n    split: train\n    type: user_defined.default\n\n    field_prompt: \"prompt\"\n    field_system: \"system\"\n    field_completion: \"completion\"\n    field_label: \"label\"\n    prompt_format: \"{prompt}\"\n    completion_format: \"{completion}\"\nThe input format is a simple JSON input with customizable fields based on the above config.\n{\n    \"system\": \"...\",  // optional\n    \"prompt\": \"...\",\n    \"completion\": \"...\",\n    \"label\": \"...\"\n}\n\n\n\nGRPO\n\n\n\n\n\n\nTip\n\n\n\nCheck out our GRPO cookbook.\n\n\nIn the latest GRPO implementation, vLLM is used to significantly speedup trajectory generation during training. In this example, we’re using 4 GPUs - 2 for training, and 2 for vLLM:\n\n\n\n\n\n\nImportant\n\n\n\nMake sure you’ve installed the correct version of vLLM by including it as an extra when installing axolotl, e.g. pip install axolotl[vllm].\n\n\nbase_model: Qwen/Qwen2.5-1.5B-Instruct\n\nvllm:\n    host: 0.0.0.0\n    port: 8000\n    tensor_parallel_size: 2\n    gpu_memory_utilization: 0.85\n    dtype: auto\n    # max_model_len: # you may find it useful to set the vLLM model context length if you know this beforehand\n\nrl: grpo\ntrl:\n    use_vllm: true\n    vllm_server_host: 0.0.0.0\n    vllm_server_port: 8000\n    vllm_server_timeout: 300\nCUDA_VISIBLE_DEVICES=2,3 axolotl vllm-serve grpo.yaml\nYour vLLM instance will now attempt to spin up, and it’s time to kick off training utilizing our remaining two GPUs. In another terminal, execute:\nCUDA_VISIBLE_DEVICES=0,1 axolotl train grpo.yaml --num-processes 2\n\n\n\n\n\n\nNote\n\n\n\nDue to TRL’s implementation with vLLM, the vLLM instance must use the last N GPUs instead of the first N GPUs. This is why in the example above, we use CUDA_VISIBLE_DEVICES=2,3 for the vLLM instance.\n\n\n\nReward functions\nGRPO uses custom reward functions and transformations. Please have them ready locally.\nFor example, to load OpenAI’s GSM8K and use a random reward for completions:\n# rewards.py\nimport random\n\ndef rand_reward_func(completions, **kwargs) -> list[float]:\n    return [random.uniform(0, 1) for _ in completions]\n\ndef oai_gsm8k_transform(cfg, *args, **kwargs):\n    def transform_fn(example, tokenizer=None):\n        label = example[\"answer\"].split(\"####\")[-1].strip().replace(\",\", \"\")\n        return {\n            \"prompt\": [{\"role\": \"user\", \"content\": example[\"question\"]},],\n            \"answer\": label,\n        }\n    return transform_fn, {\"remove_columns\": [\"question\"]}\nrl: grpo\n\ntrl:\n    beta: 0.001\n    max_completion_length: 256\n    use_vllm: True\n    num_generations: 4\n    reward_funcs: [\"rewards.rand_reward_func\"]    # format: '{file_name}.{fn_name}'\n    reward_weights: [1.0]\ndatasets:\n  - path: openai/gsm8k\n    name: main\n    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'\nTo see other examples of custom reward functions, please see TRL GRPO Docs.\nTo see all configs, please see TRLConfig.\n\n\nGRPO with DAPO/Dr. GRPO loss\nThe DAPO paper and subsequently Dr. GRPO paper proposed an alternative loss function for GRPO to remediate the penalty in longer responses.\ntrl:\n  loss_type: dr_grpo\n  # Normalizes loss based on max completion length (default: 256)\n  max_completion_length:\nFor more information, see GRPO docs.\n\n\n\nSimPO\nSimPO uses CPOTrainer but with alternative loss function.\nrl: simpo\nrl_beta: 0.1  # default in CPOTrainer\ncpo_alpha: 1.0  # default in CPOTrainer\nsimpo_gamma: 0.5  # default in CPOTrainer\nThis method uses the same dataset format as DPO.\n\n\nUsing local dataset files\ndatasets:\n  - ds_type: json\n    data_files:\n      - orca_rlhf.jsonl\n    split: train\n    type: chatml.intel\n\n\nTRL auto-unwrapping for PEFT\nTRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:\n# load ref model when adapter training.\nrl_adapter_ref_model: true",
    "crumbs": [
      "How To Guides",
      "RLHF (Beta)"
    ]
  },
  {
    "objectID": "docs/qat.html",
    "href": "docs/qat.html",
    "title": "Quantization Aware Training (QAT)",
    "section": "",
    "text": "Quantization Aware Training (QAT) is a technique for improving the accuracy of models which are quantized\nby applying “fake” quantizations to the model’s weights (and optionally, activations) during training. This fake\nquantization allows for the model to adjust for noise introduced by the quantization, so when the model is eventually\nquantized, the accuracy loss is minimized. We use the quantization techniques implemented in torchao to provide\nsupport for QAT and post-training quantization (PTQ) in axolotl.\nWe recommend reviewing the excellent QAT tutorial in the torchtune library,\nand the QAT documentation in the torchao library, for more details.",
    "crumbs": [
      "How To Guides",
      "Quantization Aware Training (QAT)"
    ]
  },
  {
    "objectID": "docs/qat.html#overview",
    "href": "docs/qat.html#overview",
    "title": "Quantization Aware Training (QAT)",
    "section": "",
    "text": "Quantization Aware Training (QAT) is a technique for improving the accuracy of models which are quantized\nby applying “fake” quantizations to the model’s weights (and optionally, activations) during training. This fake\nquantization allows for the model to adjust for noise introduced by the quantization, so when the model is eventually\nquantized, the accuracy loss is minimized. We use the quantization techniques implemented in torchao to provide\nsupport for QAT and post-training quantization (PTQ) in axolotl.\nWe recommend reviewing the excellent QAT tutorial in the torchtune library,\nand the QAT documentation in the torchao library, for more details.",
    "crumbs": [
      "How To Guides",
      "Quantization Aware Training (QAT)"
    ]
  },
  {
    "objectID": "docs/qat.html#configuring-qat-in-axolotl",
    "href": "docs/qat.html#configuring-qat-in-axolotl",
    "title": "Quantization Aware Training (QAT)",
    "section": "Configuring QAT in Axolotl",
    "text": "Configuring QAT in Axolotl\nTo enable QAT in axolotl, add the following to your configuration file:\nqat:\n  activation_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for activation quantization. Valid options are \"int4\" and \"int8\"\n  weight_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for weight quantization. Valid options are \"int4\" and \"int8\"\n  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization\n  fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after\nOnce you have finished training, you must quantize your model by using the same quantization configuration which you used to train the model with. You can use the quantize command to do this.",
    "crumbs": [
      "How To Guides",
      "Quantization Aware Training (QAT)"
    ]
  },
  {
    "objectID": "docs/fsdp_qlora.html",
    "href": "docs/fsdp_qlora.html",
    "title": "FDSP + QLoRA",
    "section": "",
    "text": "Using FSDP with QLoRA is essential for fine-tuning larger (70b+ parameter) LLMs on consumer GPUs. For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs1.\nBelow, we describe how to use this feature in Axolotl.",
    "crumbs": [
      "Advanced Features",
      "FDSP + QLoRA"
    ]
  },
  {
    "objectID": "docs/fsdp_qlora.html#background",
    "href": "docs/fsdp_qlora.html#background",
    "title": "FDSP + QLoRA",
    "section": "",
    "text": "Using FSDP with QLoRA is essential for fine-tuning larger (70b+ parameter) LLMs on consumer GPUs. For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs1.\nBelow, we describe how to use this feature in Axolotl.",
    "crumbs": [
      "Advanced Features",
      "FDSP + QLoRA"
    ]
  },
  {
    "objectID": "docs/fsdp_qlora.html#usage",
    "href": "docs/fsdp_qlora.html#usage",
    "title": "FDSP + QLoRA",
    "section": "Usage",
    "text": "Usage\nTo enable QLoRA with FSDP, you need to perform the following steps:\n\n![Tip]\nSee the example config file in addition to reading these instructions.\n\n\nSet adapter: qlora in your axolotl config file.\nEnable FSDP in your axolotl config, as described here.\nUse one of the supported model types: llama, mistral or mixtral.",
    "crumbs": [
      "Advanced Features",
      "FDSP + QLoRA"
    ]
  },
  {
    "objectID": "docs/fsdp_qlora.html#example-config",
    "href": "docs/fsdp_qlora.html#example-config",
    "title": "FDSP + QLoRA",
    "section": "Example Config",
    "text": "Example Config\nexamples/llama-2/qlora-fsdp.yml contains an example of how to enable QLoRA + FSDP in axolotl.",
    "crumbs": [
      "Advanced Features",
      "FDSP + QLoRA"
    ]
  },
  {
    "objectID": "docs/fsdp_qlora.html#references",
    "href": "docs/fsdp_qlora.html#references",
    "title": "FDSP + QLoRA",
    "section": "References",
    "text": "References\n\nPR #1378 enabling QLoRA in FSDP in Axolotl.\nBlog Post from the Answer.AI team describing the work that enabled QLoRA in FSDP.\nRelated HuggingFace PRs Enabling FDSP + QLoRA:\n\nAccelerate PR#2544\nTransformers PR#29587\nTRL PR#1416\nPEFT PR#1550",
    "crumbs": [
      "Advanced Features",
      "FDSP + QLoRA"
    ]
  },
  {
    "objectID": "docs/fsdp_qlora.html#footnotes",
    "href": "docs/fsdp_qlora.html#footnotes",
    "title": "FDSP + QLoRA",
    "section": "Footnotes",
    "text": "Footnotes\n\n\nThis was enabled by this work from the Answer.AI team.↩︎",
    "crumbs": [
      "Advanced Features",
      "FDSP + QLoRA"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html",
    "href": "docs/custom_integrations.html",
    "title": "Custom Integrations",
    "section": "",
    "text": "Axolotl adds custom features through integrations. They are located within the src/axolotl/integrations directory.\nTo enable them, please check the respective documentations.",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#cut-cross-entropy",
    "href": "docs/custom_integrations.html#cut-cross-entropy",
    "title": "Custom Integrations",
    "section": "Cut Cross Entropy",
    "text": "Cut Cross Entropy\nCut Cross Entropy (CCE) reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nRequirements\n\nPyTorch 2.4.0 or higher\n\n\n\nInstallation\nRun the following command to install cut_cross_entropy[transformers] if you don’t have it already.\n\nIf you are in dev environment\n\npython scripts/cutcrossentropy_install.py | sh\n\nIf you are installing from pip\n\npip3 uninstall -y cut-cross-entropy && pip3 install \"cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438\"\n\n\nUsage\nplugins:\n  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n\nSupported Models\n\nllama\nllama4\nllama4_text\nmllama\nphi3\ngemma\ngemma2\ngemma3\ngemma3_text\nmistral\nmistral3\nqwen2\nqwen2_moe\nqwen2_vl\nqwen2_5_vl\nqwen3\nqwen3_moe\ncohere\ncohere2\nglm\nglm4\n\n\n\nCitation\n@article{wijmans2024cut,\n  author       = {Erik Wijmans and\n                  Brody Huval and\n                  Alexander Hertzberg and\n                  Vladlen Koltun and\n                  Philipp Kr\\\"ahenb\\\"uhl},\n  title        = {Cut Your Losses in Large-Vocabulary Language Models},\n  journal      = {arXiv},\n  year         = {2024},\n  url          = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#grokfast",
    "href": "docs/custom_integrations.html#grokfast",
    "title": "Custom Integrations",
    "section": "Grokfast",
    "text": "Grokfast\nSee https://github.com/ironjr/grokfast\n\nUsage\nplugins:\n  - axolotl.integrations.grokfast.GrokfastPlugin\n\ngrokfast_alpha: 2.0\ngrokfast_lamb: 0.98\n\n\nCitation\n@article{lee2024grokfast,\n    title={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},\n    author={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},\n    journal={arXiv preprint arXiv:2405.20233},\n    year={2024}\n}\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#knowledge-distillation-kd",
    "href": "docs/custom_integrations.html#knowledge-distillation-kd",
    "title": "Custom Integrations",
    "section": "Knowledge Distillation (KD)",
    "text": "Knowledge Distillation (KD)\n\nUsage\nplugins:\n  - \"axolotl.integrations.kd.KDPlugin\"\n\nkd_trainer: True\nkd_ce_alpha: 0.1\nkd_alpha: 0.9\nkd_temperature: 1.0\n\ntorch_compile: True  # torch>=2.5.1, recommended to reduce vram\n\ndatasets:\n  - path: ...\n    type: \"axolotl.integrations.kd.chat_template\"\n    field_messages: \"messages_combined\"\n    logprobs_field: \"llm_text_generation_vllm_logprobs\"  # for kd only, field of logprobs\nAn example dataset can be found at axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#liger-kernels",
    "href": "docs/custom_integrations.html#liger-kernels",
    "title": "Custom Integrations",
    "section": "Liger Kernels",
    "text": "Liger Kernels\nLiger Kernel provides efficient Triton kernels for LLM training, offering:\n\n20% increase in multi-GPU training throughput\n60% reduction in memory usage\nCompatibility with both FSDP and DeepSpeed\n\nSee https://github.com/linkedin/Liger-Kernel\n\nUsage\nplugins:\n  - axolotl.integrations.liger.LigerPlugin\nliger_rope: true\nliger_rms_norm: true\nliger_glu_activation: true\nliger_layer_norm: true\nliger_fused_linear_cross_entropy: true\n\n\nSupported Models\n\ndeepseek_v2\ngemma\ngemma2\ngemma3\ngranite\njamba\nllama\nmistral\nmixtral\nmllama\nmllama_text_model\nolmo2\npaligemma\nphi3\nqwen2\nqwen2_5_vl\nqwen2_vl\n\n\n\nCitation\n@article{hsu2024ligerkernelefficienttriton,\n      title={Liger Kernel: Efficient Triton Kernels for LLM Training},\n      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},\n      year={2024},\n      eprint={2410.10989},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2410.10989},\n      journal={arXiv preprint arXiv:2410.10989},\n}\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#language-model-evaluation-harness-lm-eval",
    "href": "docs/custom_integrations.html#language-model-evaluation-harness-lm-eval",
    "title": "Custom Integrations",
    "section": "Language Model Evaluation Harness (LM Eval)",
    "text": "Language Model Evaluation Harness (LM Eval)\nRun evaluation on model using the popular lm-evaluation-harness library.\nSee https://github.com/EleutherAI/lm-evaluation-harness\n\nUsage\nplugins:\n  - axolotl.integrations.lm_eval.LMEvalPlugin\n\nlm_eval_tasks:\n  - gsm8k\n  - hellaswag\n  - arc_easy\n\nlm_eval_batch_size: # Batch size for evaluation\noutput_dir: # Directory to save evaluation results\n\n\nCitation\n@misc{eval-harness,\n  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},\n  title        = {A framework for few-shot language model evaluation},\n  month        = 07,\n  year         = 2024,\n  publisher    = {Zenodo},\n  version      = {v0.4.3},\n  doi          = {10.5281/zenodo.12608602},\n  url          = {https://zenodo.org/records/12608602}\n}\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#spectrum",
    "href": "docs/custom_integrations.html#spectrum",
    "title": "Custom Integrations",
    "section": "Spectrum",
    "text": "Spectrum\nby Eric Hartford, Lucas Atkins, Fernando Fernandes, David Golchinfar\nThis plugin contains code to freeze the bottom fraction of modules in a model, based on the Signal-to-Noise Ratio (SNR).\nSee https://github.com/cognitivecomputations/spectrum\n\nOverview\nSpectrum is a tool for scanning and evaluating the Signal-to-Noise Ratio (SNR) of layers in large language models.\nBy identifying the top n% of layers with the highest SNR, you can optimize training efficiency.\n\n\nUsage\nplugins:\n  - axolotl.integrations.spectrum.SpectrumPlugin\n\nspectrum_top_fraction: 0.5\nspectrum_model_name: meta-llama/Meta-Llama-3.1-8B\n\n\nCitation\n@misc{hartford2024spectrumtargetedtrainingsignal,\n      title={Spectrum: Targeted Training on Signal to Noise Ratio},\n      author={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},\n      year={2024},\n      eprint={2406.06623},\n      archivePrefix={arXiv},\n      primaryClass={cs.LG},\n      url={https://arxiv.org/abs/2406.06623},\n}\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#llmcompressor",
    "href": "docs/custom_integrations.html#llmcompressor",
    "title": "Custom Integrations",
    "section": "LLMCompressor",
    "text": "LLMCompressor\nFine-tune sparsified models in Axolotl using Neural Magic’s LLMCompressor.\nThis integration enables fine-tuning of models sparsified using LLMCompressor within the Axolotl training framework. By combining LLMCompressor’s model compression capabilities with Axolotl’s distributed training pipelines, users can efficiently fine-tune sparse models at scale.\nIt uses Axolotl’s plugin system to hook into the fine-tuning flows while maintaining sparsity throughout training.\n\n\nRequirements\n\nAxolotl with llmcompressor extras:\npip install \"axolotl[llmcompressor]\"\nRequires llmcompressor >= 0.5.1\n\nThis will install all necessary dependencies to fine-tune sparsified models using the integration.\n\n\n\nUsage\nTo enable sparse fine-tuning with this integration, include the plugin in your Axolotl config:\nplugins:\n  - axolotl.integrations.llm_compressor.LLMCompressorPlugin\n\nllmcompressor:\n  recipe:\n    finetuning_stage:\n      finetuning_modifiers:\n        ConstantPruningModifier:\n          targets: [\n            're:.*q_proj.weight',\n            're:.*k_proj.weight',\n            're:.*v_proj.weight',\n            're:.*o_proj.weight',\n            're:.*gate_proj.weight',\n            're:.*up_proj.weight',\n            're:.*down_proj.weight',\n          ]\n          start: 0\n  save_compressed: true\nThis plugin does not apply pruning or sparsification itself — it is intended for fine-tuning models that have already been sparsified.\nPre-sparsified checkpoints can be:\n- Generated using LLMCompressor\n- Downloaded from Neural Magic’s Hugging Face page\n- Any custom LLM with compatible sparsity patterns that you’ve created yourself\nTo learn more about writing and customizing LLMCompressor recipes, refer to the official documentation:\nhttps://github.com/vllm-project/llm-compressor/blob/main/README.md\n\n\nStorage Optimization with save_compressed\nSetting save_compressed: true in your configuration enables saving models in a compressed format, which:\n- Reduces disk space usage by approximately 40%\n- Maintains compatibility with vLLM for accelerated inference\n- Maintains compatibility with llmcompressor for further optimization (example: quantization)\nThis option is highly recommended when working with sparse models to maximize the benefits of model compression.\n\n\nExample Config\nSee examples/llama-3/sparse-finetuning.yaml for a complete example.\n\n\n\nInference with vLLM\nAfter fine-tuning your sparse model, you can leverage vLLM for efficient inference.\nYou can also use LLMCompressor to apply additional quantization to your fine-tuned\nsparse model before inference for even greater performance benefits.:\nfrom vllm import LLM, SamplingParams\n\nprompts = [\n    \"Hello, my name is\",\n    \"The president of the United States is\",\n    \"The capital of France is\",\n    \"The future of AI is\",\n]\nsampling_params = SamplingParams(temperature=0.8, top_p=0.95)\nllm = LLM(\"path/to/your/sparse/model\")\noutputs = llm.generate(prompts, sampling_params)\n\nfor output in outputs:\n    prompt = output.prompt\n    generated_text = output.outputs[0].text\n    print(f\"Prompt: {prompt!r}, Generated text: {generated_text!r}\")\nFor more details on vLLM’s capabilities and advanced configuration options, see the official vLLM documentation.\n\n\nLearn More\nFor details on available sparsity and quantization schemes, fine-tuning recipes, and usage examples, visit the official LLMCompressor repository:\nhttps://github.com/vllm-project/llm-compressor\nPlease see reference here",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/custom_integrations.html#adding-a-new-integration",
    "href": "docs/custom_integrations.html#adding-a-new-integration",
    "title": "Custom Integrations",
    "section": "Adding a new integration",
    "text": "Adding a new integration\nPlugins can be used to customize the behavior of the training pipeline through hooks. See axolotl.integrations.BasePlugin for the possible hooks.\nTo add a new integration, please follow these steps:\n\nCreate a new folder in the src/axolotl/integrations directory.\nAdd any relevant files (LICENSE, README.md, ACKNOWLEDGEMENTS.md, etc.) to the new folder.\nAdd __init__.py and args.py files to the new folder.\n\n\n__init__.py should import the integration and hook into the appropriate functions.\nargs.py should define the arguments for the integration.\n\n\n(If applicable) Add CPU tests under tests/integrations or GPU tests under tests/e2e/integrations.\n\n\n\n\n\n\n\nTip\n\n\n\nSee src/axolotl/integrations/cut_cross_entropy for a minimal integration example.\n\n\n\n\n\n\n\n\nWarning\n\n\n\nIf you could not load your integration, please ensure you are pip installing in editable mode.\npip install -e .\nand correctly spelled the integration name in the config file.\nplugins:\n  - axolotl.integrations.your_integration_name.YourIntegrationPlugin\n\n\n\n\n\n\n\n\nNote\n\n\n\nIt is not necessary to place your integration in the integrations folder. It can be in any location, so long as it’s installed in a package in your python env.\nSee this repo for an example: https://github.com/axolotl-ai-cloud/diff-transformer",
    "crumbs": [
      "Advanced Features",
      "Custom Integrations"
    ]
  },
  {
    "objectID": "docs/dataset-formats/pretraining.html",
    "href": "docs/dataset-formats/pretraining.html",
    "title": "Pre-training",
    "section": "",
    "text": "For pretraining, there is no prompt template or roles. The only required field is text:\n\n\ndata.jsonl\n\n{\"text\": \"first row\"}\n{\"text\": \"second row\"}\n...\n\n\n\n\n\n\n\nStreaming is recommended for large datasets\n\n\n\nAxolotl usually loads the entire dataset into memory. This will be challenging for large datasets. Use the following config to enable streaming:\n\n\nconfig.yaml\n\npretraining_dataset:\n  - name:\n    path:\n    split:\n    text_column: # column in dataset with the data, usually `text`\n    type: pretrain\n    trust_remote_code:\n    skip: # number of rows of data to skip over from the beginning",
    "crumbs": [
      "Dataset Formats",
      "Pre-training"
    ]
  },
  {
    "objectID": "docs/dataset-formats/template_free.html",
    "href": "docs/dataset-formats/template_free.html",
    "title": "Template-Free",
    "section": "",
    "text": "One of the most popular features of\naxolotl is\nsetting the following configuration value:\ntrain_on_inputs: false\nIf you declare a dataset formats\nsuch as alpaca or chatml, axolotl knows what is an input\n(i.e. human) vs. an output (i.e. the assistant) and masks the input\nlabels so that your model can focus on predicting the outputs only.\n\n\n\nHowever, there are many situations where you don’t want to use one of\nthese formats or templates. This is because they can:\n\nAdd unnecessary boilerplate to your prompts.\nCreate artifacts like special delimiters <|im_start|> that can\nquickly become footguns if you don’t include them correctly at\ninference time.\nEnforce a chat interface when you do not want one. Sometimes you\njust want to fine-tune a model to a very specific task and do NOT\nwant multi-turn conversations, roles, etc.\nLimit you to only certain roles that the template allows.\n\n\n\n\nYou can construct your prompts without a template by using the\ninput_output format, by setting type: input_output in your\nconfiguration file like this:\nconfig.yml\ntrain_on_inputs: false # Mask segments of your data\ndatasets:\n  - path: output.jsonl\n    type: input_output  # use template free prompt construction\nUnlike type: completion, which is also template-free,\ntype: input_output allows you to mask segments of your text. More\ndetails on how this works are described below.",
    "crumbs": [
      "Dataset Formats",
      "Template-Free"
    ]
  },
  {
    "objectID": "docs/dataset-formats/template_free.html#sec-background",
    "href": "docs/dataset-formats/template_free.html#sec-background",
    "title": "Template-Free",
    "section": "",
    "text": "One of the most popular features of\naxolotl is\nsetting the following configuration value:\ntrain_on_inputs: false\nIf you declare a dataset formats\nsuch as alpaca or chatml, axolotl knows what is an input\n(i.e. human) vs. an output (i.e. the assistant) and masks the input\nlabels so that your model can focus on predicting the outputs only.\n\n\n\nHowever, there are many situations where you don’t want to use one of\nthese formats or templates. This is because they can:\n\nAdd unnecessary boilerplate to your prompts.\nCreate artifacts like special delimiters <|im_start|> that can\nquickly become footguns if you don’t include them correctly at\ninference time.\nEnforce a chat interface when you do not want one. Sometimes you\njust want to fine-tune a model to a very specific task and do NOT\nwant multi-turn conversations, roles, etc.\nLimit you to only certain roles that the template allows.\n\n\n\n\nYou can construct your prompts without a template by using the\ninput_output format, by setting type: input_output in your\nconfiguration file like this:\nconfig.yml\ntrain_on_inputs: false # Mask segments of your data\ndatasets:\n  - path: output.jsonl\n    type: input_output  # use template free prompt construction\nUnlike type: completion, which is also template-free,\ntype: input_output allows you to mask segments of your text. More\ndetails on how this works are described below.",
    "crumbs": [
      "Dataset Formats",
      "Template-Free"
    ]
  },
  {
    "objectID": "docs/dataset-formats/template_free.html#sec-usage",
    "href": "docs/dataset-formats/template_free.html#sec-usage",
    "title": "Template-Free",
    "section": "Usage",
    "text": "Usage\nThis is how you can use the input_output format:\n\n1. Prepare Data\nTo use the input_output format, collect your data in the following\nformat into a jsonl file (below is the first row from the file\noutput.jsonl` pretty printed):\n$ head -n1 output.jsonl | python -m json.tool\n\n{\n    \"segments\": [\n        {\n            \"label\": true,\n            \"text\": \"<s>Hello\\n\"\n        },\n        {\n            \"label\": true,\n            \"text\": \"hi there!. \"\n        },\n        {\n            \"label\": false,\n            \"text\": \"goodbye \"\n        },\n        {\n            \"label\": true,\n            \"text\": \"farewell</s>\"\n        }\n    ]\n}\n\nSet label:false when you want to mask a segment of text so that the\nmodel isn’t trained on it. Some things to keep in mind:\n\n[!IMPORTANT]\n1. EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl\nconcatenates all the segments as-is. The tokenizer doesn’t add\nanything additional. Notice how I added spaces, newlines, <s>\n(BOS), and </s> (EOS) myself.\n2. Make sure you check the materialized output to validate that the\nprompt is getting assembled how you like.\n\n\n\n2. Use type: input_output\nLet’s materialize data with our output.jsonl file by setting\ntype: input_output in our axolotl config:\n# training_config.yaml\nbase_model: mistralai/Mistral-7B-v0.1\ndata_seed: 49\nseed: 49\n\ndatasets:\n  - path: output.jsonl\n    type: input_output\nval_set_size: 0.1\n\nsequence_len: 896\nsample_packing: false\n\nmicro_batch_size: 2\ngradient_accumulation_steps: 3\neval_batch_size: 2\nnum_epochs: 1\nlearning_rate: 0.0002\n\ntrain_on_inputs: false\nspecial_tokens:\n  bos_token: \"<s>\"\n  eos_token: \"</s>\"\n  unk_token: \"<unk>\"\nYou can use the following command to materialize your data. The\n--debug flag will print the tokens, along with the labels so you can\nverify that the correct items are being ignored:\naxolotl preprocess training_config.yaml --debug\n\n...\n[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)\n(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)\nThe format is decoded_token(label, token_id), for example,\n<s>(1, 1) means that the token is <s>, the label is 1 and the\ntoken_id is 1. When the label is -100 then that token is ignored for\ntraining.\n\n\n3. Check the prompts\nHere is another way to check the materialized output:\nfrom transformers import AutoTokenizer\nfrom datasets import load_from_disk\nimport yaml\n\ndirectory = !ls last_run_prepared/\nwith open('training_config.yaml', 'r') as f:\n    cfg = yaml.safe_load(f)\nmodel_id = cfg['base_model']\ntok = AutoTokenizer.from_pretrained(model_id)\nds = load_from_disk(f'last_run_prepared/{directory[0]}/')\n>>> row = ds[0]\n>>> print(tok.decode(row['input_ids']))\n<s> Hello\n    hi there!.  goodbye  farewell</s>\nWe can check that the right tokens are ignored by comparing the labels\nto each token:\nimport pandas as pd\npd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in\n              zip(row['input_ids'], row['labels'])])\n\n\n\ntoken\nlabel\nid\n\n\n\n\n0\n<s>\n1\n\n\n1\nHello\n22557\n\n\n2\n\\n\n13\n\n\n3\nhi\n12014\n\n\n4\nthere\n736\n\n\n5\n!\n28808\n\n\n6\n.\n28723\n\n\n7\n\n28705\n\n\n8\ngood\n-100\n\n\n9\nbye\n-100\n\n\n10\n\n-100\n\n\n11\nfare\n19111\n\n\n12\nwell\n5458\n\n\n13\n</s>\n2\n\n\n\nIf we look at the input data, the above table seems correct! (The jsonl\nversion is repeated below for reference):\n$ head -n1 output.jsonl | python -m json.tool\n\n{\n    \"segments\": [\n        {\n            \"label\": true,\n            \"text\": \"<s>Hello\\n\"\n        },\n        {\n            \"label\": true,\n            \"text\": \"hi there!. \"\n        },\n        {\n            \"label\": false,\n            \"text\": \"goodbye \"\n        },\n        {\n            \"label\": true,\n            \"text\": \"farewell</s>\"\n        }\n    ]\n}",
    "crumbs": [
      "Dataset Formats",
      "Template-Free"
    ]
  },
  {
    "objectID": "docs/dataset-formats/conversation.html",
    "href": "docs/dataset-formats/conversation.html",
    "title": "Conversation",
    "section": "",
    "text": "Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer’s template, a supported template, or custom jinja2.\n\n\ndata.jsonl\n\n{\"conversations\": [{\"role\": \"...\", \"content\": \"...\"}]}\n\nSee configs for full configs and supported templates.\n\n\nMost configs can be adapted as follows:\n# old\nchat_template: chatml\ndatasets:\n  - path: ...\n    type: sharegpt\n    conversation: chatml\n\n# new (if using tokenizer's chat_template)\ndatasets:\n  - path: ...\n    type: chat_template\n\n    field_messages: conversations\n    message_property_mappings:\n      role: from\n      content: value\n\n# new (if setting a new chat_template like chatml, gemma, etc)\nchat_template: chatml\ndatasets:\n  - path: ...\n    type: chat_template\n\n    field_messages: conversations\n    message_property_mappings:\n      role: from\n      content: value\nWe recommend checking the below examples for other usecases.\n\n\n\n\n(Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.\n\ndatasets:\n  - path: ...\n    type: chat_template\n    roles_to_train:\n    train_on_eos:\n\n\n\n\n\n\nTip\n\n\n\nIf you receive an error like “chat_template choice is tokenizer_default but tokenizer’s chat_template is null.”, it means the tokenizer does not have a default chat_template. Follow the examples below instead to set a custom chat_template.\n\n\n\nUsing the gemma chat template to override the tokenizer_config.json’s chat template on OpenAI messages format, training on all assistant messages.\n\nchat_template: gemma # this overwrites the tokenizer's chat_template\ndatasets:\n  - path: ...\n    type: chat_template\n    roles_to_train: [\"assistant\"]  # default value\n\nUsing the tokenizer_config.json’s chat template or chatml as fallback if the former’s chat template does not exist, on OpenAI messages format, training on all assistant messages.\n\nchat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template\ndatasets:\n  - path: ...\n    type: chat_template\n\nUsing a custom jinja template on OpenAI messages format, training on all assistant messages.\n\n# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty\nchat_template_jinja: \"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\\n' + message['content'] + '<|end|>' + '\\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\\n' + message['content'] + '<|end|>' + '\\n' + '<|assistant|>' + '\\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\\n'}}{% endif %}{% endfor %}\"\n\ndatasets:\n  - path: ...\n    type: chat_template\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure that your tokenizer.eos_token is same as EOS (End-of-Sequence) token in template. Otherwise, set eos_token under special_tokens:.\n\n\n\nIf you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the eot_tokens: config. The handling of EOT tokens follows train_on_eos: which defaults to turn.\n\neot_tokens:\n  - \"[/INST]\"\n  # - \"[/SYSTEM_PROMPT]\"\n\ndatasets:\n  - path: ...\n    type: chat_template\n\n    # optional\n    train_on_eot: turn  # defaults read from train_on_eos (which defaults to turn)\n\n\n\n\n\n\nTip\n\n\n\nSee config documentation for detailed explanations of “turn”, “last”, and “all” options for training on tokens.\n\n\n\n\n\n\n\n\nNote\n\n\n\nUsing eot_tokens requires each token that exists in chat_template to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.\nYou can add those tokens as new tokens under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:. See config for more details.\n\n\n\nContinuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set train_on_eos: last.\n\neot_tokens:\n  - \"[/INST]\"\n  # ...\n\ndatasets:\n  - path: ...\n    type: chat_template\n\n    train_on_eos: last\n    train_on_eot: turn\n\n\n\n\n\n\nTip\n\n\n\nIf EOS token only appears at the end of a prompt, train_on_eos: last is equivalent to train_on_eos: turn. Therefore, generally, you can leave them to their defaults and omit them.\n\n\n\n(Advanced) Using fine-grained control over tokens and turns to train in a conversation\n\nFor a data sample that looks like:\n\n\ndata.jsonl\n\n{\n  \"conversations\": [\n    {\"from\": \"system\", \"value\": \"You are an AI assistant.\", \"train\": false},\n    {\"from\": \"human\", \"value\": \"Hello\", \"train\": false},\n    {\"from\": \"assistant\", \"value\": \"Hello\", \"train\": true},\n    {\"from\": \"human\", \"value\": \"How are you?\", \"train\": true},\n    {\n      \"from\": \"assistant\",\n      \"value\": \"I'm doing very well, thank you!\",\n      \"train_detail\": [\n        {\"begin_offset\": 0, \"end_offset\": 8, \"train\": false},\n        {\"begin_offset\": 9, \"end_offset\": 18, \"train\": true},\n        {\"begin_offset\": 19, \"end_offset\": 30, \"train\": false},\n      ],\n    },\n    {\n        \"from\": \"human\",\n        \"value\": \"I'm doing very well, thank you!\",\n        \"train\": true,\n    },\n    {\"from\": \"assistant\", \"value\": \"Hi there!\", \"train\": true}\n  ]\n}\n\nThe configuration would look like:\ndatasets:\n  - path: ...\n    type: chat_template\n    chat_template: tokenizer_default\n    field_messages: conversations\n    message_property_mappings:\n      role: from\n      content: value\n    roles_to_train: []\n    train_on_eos: turn\n    message_field_training: train\n    message_field_training_detail: train_detail\n\n\n\n\n\n\nTip\n\n\n\nIt is not necessary to set both message_field_training and message_field_training_detail at once.\n\n\n\n(For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.\n\ndatasets:\n  - path: ...\n    type: chat_template\n    chat_template: qwen3\n    split_thinking: true\nFor example, a content can look like:\n{\n  \"content\": \"<think>Some thinking outputs</think>Output after thinking.\"\n}\nAfter split, it will look like:\n{\n  \"reasoning_content\": \"Some thinking outputs\",\n  \"content\": \"Output after thinking...\"\n}",
    "crumbs": [
      "Dataset Formats",
      "Conversation"
    ]
  },
  {
    "objectID": "docs/dataset-formats/conversation.html#chat_template",
    "href": "docs/dataset-formats/conversation.html#chat_template",
    "title": "Conversation",
    "section": "",
    "text": "Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer’s template, a supported template, or custom jinja2.\n\n\ndata.jsonl\n\n{\"conversations\": [{\"role\": \"...\", \"content\": \"...\"}]}\n\nSee configs for full configs and supported templates.\n\n\nMost configs can be adapted as follows:\n# old\nchat_template: chatml\ndatasets:\n  - path: ...\n    type: sharegpt\n    conversation: chatml\n\n# new (if using tokenizer's chat_template)\ndatasets:\n  - path: ...\n    type: chat_template\n\n    field_messages: conversations\n    message_property_mappings:\n      role: from\n      content: value\n\n# new (if setting a new chat_template like chatml, gemma, etc)\nchat_template: chatml\ndatasets:\n  - path: ...\n    type: chat_template\n\n    field_messages: conversations\n    message_property_mappings:\n      role: from\n      content: value\nWe recommend checking the below examples for other usecases.\n\n\n\n\n(Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.\n\ndatasets:\n  - path: ...\n    type: chat_template\n    roles_to_train:\n    train_on_eos:\n\n\n\n\n\n\nTip\n\n\n\nIf you receive an error like “chat_template choice is tokenizer_default but tokenizer’s chat_template is null.”, it means the tokenizer does not have a default chat_template. Follow the examples below instead to set a custom chat_template.\n\n\n\nUsing the gemma chat template to override the tokenizer_config.json’s chat template on OpenAI messages format, training on all assistant messages.\n\nchat_template: gemma # this overwrites the tokenizer's chat_template\ndatasets:\n  - path: ...\n    type: chat_template\n    roles_to_train: [\"assistant\"]  # default value\n\nUsing the tokenizer_config.json’s chat template or chatml as fallback if the former’s chat template does not exist, on OpenAI messages format, training on all assistant messages.\n\nchat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template\ndatasets:\n  - path: ...\n    type: chat_template\n\nUsing a custom jinja template on OpenAI messages format, training on all assistant messages.\n\n# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty\nchat_template_jinja: \"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\\n' + message['content'] + '<|end|>' + '\\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\\n' + message['content'] + '<|end|>' + '\\n' + '<|assistant|>' + '\\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\\n'}}{% endif %}{% endfor %}\"\n\ndatasets:\n  - path: ...\n    type: chat_template\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure that your tokenizer.eos_token is same as EOS (End-of-Sequence) token in template. Otherwise, set eos_token under special_tokens:.\n\n\n\nIf you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the eot_tokens: config. The handling of EOT tokens follows train_on_eos: which defaults to turn.\n\neot_tokens:\n  - \"[/INST]\"\n  # - \"[/SYSTEM_PROMPT]\"\n\ndatasets:\n  - path: ...\n    type: chat_template\n\n    # optional\n    train_on_eot: turn  # defaults read from train_on_eos (which defaults to turn)\n\n\n\n\n\n\nTip\n\n\n\nSee config documentation for detailed explanations of “turn”, “last”, and “all” options for training on tokens.\n\n\n\n\n\n\n\n\nNote\n\n\n\nUsing eot_tokens requires each token that exists in chat_template to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.\nYou can add those tokens as new tokens under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:. See config for more details.\n\n\n\nContinuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set train_on_eos: last.\n\neot_tokens:\n  - \"[/INST]\"\n  # ...\n\ndatasets:\n  - path: ...\n    type: chat_template\n\n    train_on_eos: last\n    train_on_eot: turn\n\n\n\n\n\n\nTip\n\n\n\nIf EOS token only appears at the end of a prompt, train_on_eos: last is equivalent to train_on_eos: turn. Therefore, generally, you can leave them to their defaults and omit them.\n\n\n\n(Advanced) Using fine-grained control over tokens and turns to train in a conversation\n\nFor a data sample that looks like:\n\n\ndata.jsonl\n\n{\n  \"conversations\": [\n    {\"from\": \"system\", \"value\": \"You are an AI assistant.\", \"train\": false},\n    {\"from\": \"human\", \"value\": \"Hello\", \"train\": false},\n    {\"from\": \"assistant\", \"value\": \"Hello\", \"train\": true},\n    {\"from\": \"human\", \"value\": \"How are you?\", \"train\": true},\n    {\n      \"from\": \"assistant\",\n      \"value\": \"I'm doing very well, thank you!\",\n      \"train_detail\": [\n        {\"begin_offset\": 0, \"end_offset\": 8, \"train\": false},\n        {\"begin_offset\": 9, \"end_offset\": 18, \"train\": true},\n        {\"begin_offset\": 19, \"end_offset\": 30, \"train\": false},\n      ],\n    },\n    {\n        \"from\": \"human\",\n        \"value\": \"I'm doing very well, thank you!\",\n        \"train\": true,\n    },\n    {\"from\": \"assistant\", \"value\": \"Hi there!\", \"train\": true}\n  ]\n}\n\nThe configuration would look like:\ndatasets:\n  - path: ...\n    type: chat_template\n    chat_template: tokenizer_default\n    field_messages: conversations\n    message_property_mappings:\n      role: from\n      content: value\n    roles_to_train: []\n    train_on_eos: turn\n    message_field_training: train\n    message_field_training_detail: train_detail\n\n\n\n\n\n\nTip\n\n\n\nIt is not necessary to set both message_field_training and message_field_training_detail at once.\n\n\n\n(For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.\n\ndatasets:\n  - path: ...\n    type: chat_template\n    chat_template: qwen3\n    split_thinking: true\nFor example, a content can look like:\n{\n  \"content\": \"<think>Some thinking outputs</think>Output after thinking.\"\n}\nAfter split, it will look like:\n{\n  \"reasoning_content\": \"Some thinking outputs\",\n  \"content\": \"Output after thinking...\"\n}",
    "crumbs": [
      "Dataset Formats",
      "Conversation"
    ]
  },
  {
    "objectID": "docs/dataset-formats/conversation.html#sharegpt",
    "href": "docs/dataset-formats/conversation.html#sharegpt",
    "title": "Conversation",
    "section": "sharegpt",
    "text": "sharegpt\n\n\n\n\n\n\nImportant\n\n\n\nShareGPT is deprecated!. Please see chat_template section.",
    "crumbs": [
      "Dataset Formats",
      "Conversation"
    ]
  },
  {
    "objectID": "docs/dataset-formats/conversation.html#pygmalion",
    "href": "docs/dataset-formats/conversation.html#pygmalion",
    "title": "Conversation",
    "section": "pygmalion",
    "text": "pygmalion\n\n\ndata.jsonl\n\n{\"conversations\": [{\"role\": \"...\", \"value\": \"...\"}]}",
    "crumbs": [
      "Dataset Formats",
      "Conversation"
    ]
  },
  {
    "objectID": "docs/dataset-formats/tokenized.html",
    "href": "docs/dataset-formats/tokenized.html",
    "title": "Custom Pre-Tokenized Dataset",
    "section": "",
    "text": "Pass an empty type: in your axolotl config.\nColumns in Dataset must be exactly input_ids, attention_mask, labels\nTo indicate that a token should be ignored during training, set its corresponding label to -100.\nYou must add BOS and EOS, and make sure that you are training on EOS by not setting its label to -100.\nFor pretraining, do not truncate/pad documents to the context window length.\nFor instruction training, documents must be truncated/padded as desired.\n\nSample config:\n\n\nconfig.yml\n\ndatasets:\n  - path: /path/to/your/file.jsonl\n    ds_type: json\n    type:\n\nSample jsonl:\n{\"input_ids\":[271,299,99],\"attention_mask\":[1,1,1],\"labels\":[271,-100,99]}\n{\"input_ids\":[87,227,8383,12],\"attention_mask\":[1,1,1,1],\"labels\":[87,227,8383,12]}",
    "crumbs": [
      "Dataset Formats",
      "Custom Pre-Tokenized Dataset"
    ]
  },
  {
    "objectID": "docs/multipack.html",
    "href": "docs/multipack.html",
    "title": "Multipack (Sample Packing)",
    "section": "",
    "text": "Because Flash Attention simply drops the attention mask, we do not need to\nconstruct a 4d attention mask. We only need to concatenate the sequences into\na single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4,\neach character represents 256 tokens\nX represents a padding token\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B ]\n   C C C C C C C ]\n   D D D D ]]\n\n[[ E E E E E E E E ]\n [ F F F F ]\n [ G G G ]\n [ H H H H ]]\n\n[[ I I I ]\n [ J J J ]\n [ K K K K K]\n [ L L L ]]\nafter padding to longest input in each step\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B X X X X X X ]\n   C C C C C C C X X X X ]\n   D D D D X X X X X X X ]]\n\n[[ E E E E E E E E ]\n [ F F F F X X X X ]\n [ G G G X X X X X ]\n [ H H H H X X X X ]]\n\n[[ I I I X X ]\n [ J J J X X ]\n [ K K K K K ]\n [ L L L X X ]]\nw packing ( note it’s the same effective number of tokens per step, but a true bsz of 1)\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A B B B B B\n   B C C C C C C C D D D D E E E E\n   E E E E F F F F F G G G H H H H\n   I I I J J J J K K K K K L L L X ]]\ncu_seqlens:\n[[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
    "crumbs": [
      "Core Concepts",
      "Multipack (Sample Packing)"
    ]
  },
  {
    "objectID": "docs/multipack.html#visualization-of-multipack-with-flash-attention",
    "href": "docs/multipack.html#visualization-of-multipack-with-flash-attention",
    "title": "Multipack (Sample Packing)",
    "section": "",
    "text": "Because Flash Attention simply drops the attention mask, we do not need to\nconstruct a 4d attention mask. We only need to concatenate the sequences into\na single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4,\neach character represents 256 tokens\nX represents a padding token\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B ]\n   C C C C C C C ]\n   D D D D ]]\n\n[[ E E E E E E E E ]\n [ F F F F ]\n [ G G G ]\n [ H H H H ]]\n\n[[ I I I ]\n [ J J J ]\n [ K K K K K]\n [ L L L ]]\nafter padding to longest input in each step\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A ]\n   B B B B B B X X X X X X ]\n   C C C C C C C X X X X ]\n   D D D D X X X X X X X ]]\n\n[[ E E E E E E E E ]\n [ F F F F X X X X ]\n [ G G G X X X X X ]\n [ H H H H X X X X ]]\n\n[[ I I I X X ]\n [ J J J X X ]\n [ K K K K K ]\n [ L L L X X ]]\nw packing ( note it’s the same effective number of tokens per step, but a true bsz of 1)\n   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5\n[[ A A A A A A A A A A A B B B B B\n   B C C C C C C C D D D D E E E E\n   E E E E F F F F F G G G H H H H\n   I I I J J J J K K K K K L L L X ]]\ncu_seqlens:\n[[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
    "crumbs": [
      "Core Concepts",
      "Multipack (Sample Packing)"
    ]
  },
  {
    "objectID": "docs/multipack.html#multipack-without-flash-attention",
    "href": "docs/multipack.html#multipack-without-flash-attention",
    "title": "Multipack (Sample Packing)",
    "section": "Multipack without Flash Attention",
    "text": "Multipack without Flash Attention\nMultipack can still be achieved without Flash attention, but with lower packing\nefficiency as we are not able to join multiple batches into a single batch due to\ncontext length limits without flash attention. We can use either Pytorch’s Scaled\nDot Product Attention implementation or native Pytorch attention implementation\nalong with 4d attention masks\nto pack sequences together and avoid cross attention.",
    "crumbs": [
      "Core Concepts",
      "Multipack (Sample Packing)"
    ]
  },
  {
    "objectID": "docs/mac.html",
    "href": "docs/mac.html",
    "title": "Mac M-series",
    "section": "",
    "text": "Currently Axolotl on Mac is partially usable, many of the dependencies of Axolotl including Pytorch do not support MPS or have incomplete support.\nCurrent support:\n\nSupport for all models\nFull training of models\nLoRA training\nSample packing\nFP16 and BF16 (awaiting AMP support for MPS in Pytorch)\nTri-dao’s flash-attn (until it is supported use spd_attention as an alternative)\nxformers\nbitsandbytes (meaning no 4/8 bits loading and bnb optimizers)\nqlora\nDeepSpeed\n\nUntested:\n\nFSDP",
    "crumbs": [
      "Deployments",
      "Mac M-series"
    ]
  },
  {
    "objectID": "docs/input_output.html",
    "href": "docs/input_output.html",
    "title": "Template-free prompt construction",
    "section": "",
    "text": "The documentation moved to here."
  },
  {
    "objectID": "docs/quantize.html",
    "href": "docs/quantize.html",
    "title": "Quantization with torchao",
    "section": "",
    "text": "Quantization is a technique to lower the memory footprint of your model, potentially at the cost of accuracy or model performance. We support quantizing your model using the torchao library. Quantization is supported for both post-training quantization (PTQ) and quantization-aware training (QAT).",
    "crumbs": [
      "How To Guides",
      "Quantization with torchao"
    ]
  },
  {
    "objectID": "docs/quantize.html#configuring-quantization-in-axolotl",
    "href": "docs/quantize.html#configuring-quantization-in-axolotl",
    "title": "Quantization with torchao",
    "section": "Configuring Quantization in Axolotl",
    "text": "Configuring Quantization in Axolotl\nQuantization is configured using the quantization key in your configuration file.\nbase_model: # The path to the model to quantize.\nquantization:\n  weight_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8\n  activation_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for activation quantization. Valid options are \"int4\" and \"int8\"\n  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization\n  quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer.\n\noutput_dir:  # The path to the output directory.\nOnce quantization is complete, your quantized model will be saved in the {output_dir}/quantized directory.\nYou may also use the quantize command to quantize a model which has been trained with QAT - you can do this by using the existing QAT configuration file which\nyou used to train the model:\n# qat.yml\nqat:\n  activation_dtype: int8\n  weight_dtype: int8\n  group_size: 256\n  quantize_embedding: true\n\noutput_dir: # The path to the output directory used during training where the final checkpoint has been saved.\naxolotl quantize qat.yml\nThis ensures that an identical quantization configuration is used to quantize the model as was used to train it.",
    "crumbs": [
      "How To Guides",
      "Quantization with torchao"
    ]
  },
  {
    "objectID": "docs/cli.html",
    "href": "docs/cli.html",
    "title": "Command Line Interface (CLI)",
    "section": "",
    "text": "The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers\nthe CLI commands, their usage, and common examples.",
    "crumbs": [
      "Getting Started",
      "Command Line Interface (CLI)"
    ]
  },
  {
    "objectID": "docs/cli.html#basic-commands",
    "href": "docs/cli.html#basic-commands",
    "title": "Command Line Interface (CLI)",
    "section": "Basic Commands",
    "text": "Basic Commands\nAll Axolotl commands follow this general structure:\naxolotl <command> [config.yml] [options]\nThe config file can be local or a URL to a raw YAML file.",
    "crumbs": [
      "Getting Started",
      "Command Line Interface (CLI)"
    ]
  },
  {
    "objectID": "docs/cli.html#command-reference",
    "href": "docs/cli.html#command-reference",
    "title": "Command Line Interface (CLI)",
    "section": "Command Reference",
    "text": "Command Reference\n\nfetch\nDownloads example configurations and deepspeed configs to your local machine.\n# Get example YAML files\naxolotl fetch examples\n\n# Get deepspeed config files\naxolotl fetch deepspeed_configs\n\n# Specify custom destination\naxolotl fetch examples --dest path/to/folder\n\n\npreprocess\nPreprocesses and tokenizes your dataset before training. This is recommended for large datasets.\n# Basic preprocessing\naxolotl preprocess config.yml\n\n# Preprocessing with one GPU\nCUDA_VISIBLE_DEVICES=\"0\" axolotl preprocess config.yml\n\n# Debug mode to see processed examples\naxolotl preprocess config.yml --debug\n\n# Debug with limited examples\naxolotl preprocess config.yml --debug --debug-num-examples 5\nConfiguration options:\ndataset_prepared_path: Local folder for saving preprocessed data\npush_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)\n\n\ntrain\nTrains or fine-tunes a model using the configuration specified in your YAML file.\n# Basic training\naxolotl train config.yml\n\n# Train and set/override specific options\naxolotl train config.yml \\\n    --learning-rate 1e-4 \\\n    --micro-batch-size 2 \\\n    --num-epochs 3\n\n# Training without accelerate\naxolotl train config.yml --no-accelerate\n\n# Resume training from checkpoint\naxolotl train config.yml --resume-from-checkpoint path/to/checkpoint\nIt is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.\n# Basic training with sweeps\naxolotl train config.yml --sweep path/to/sweep.yaml\nExample sweep config:\n_:\n  # This section is for dependent variables we need to fix\n  - load_in_8bit: false\n    load_in_4bit: false\n    adapter: lora\n  - load_in_8bit: true\n    load_in_4bit: false\n    adapter: lora\n\n# These are independent variables\nlearning_rate: [0.0003, 0.0006]\nlora_r:\n  - 16\n  - 32\nlora_alpha:\n  - 16\n  - 32\n  - 64\n\n\ninference\nRuns inference using your trained model in either CLI or Gradio interface mode.\n# CLI inference with LoRA\naxolotl inference config.yml --lora-model-dir=\"./outputs/lora-out\"\n\n# CLI inference with full model\naxolotl inference config.yml --base-model=\"./completed-model\"\n\n# Gradio web interface\naxolotl inference config.yml --gradio \\\n    --lora-model-dir=\"./outputs/lora-out\"\n\n# Inference with input from file\ncat prompt.txt | axolotl inference config.yml \\\n    --base-model=\"./completed-model\"\n\n\nmerge-lora\nMerges trained LoRA adapters into the base model.\n# Basic merge\naxolotl merge-lora config.yml\n\n# Specify LoRA directory (usually used with checkpoints)\naxolotl merge-lora config.yml --lora-model-dir=\"./lora-output/checkpoint-100\"\n\n# Merge using CPU (if out of GPU memory)\nCUDA_VISIBLE_DEVICES=\"\" axolotl merge-lora config.yml\nConfiguration options:\ngpu_memory_limit: Limit GPU memory usage\nlora_on_cpu: Load LoRA weights on CPU\n\n\nmerge-sharded-fsdp-weights\nMerges sharded FSDP model checkpoints into a single combined checkpoint.\n# Basic merge\naxolotl merge-sharded-fsdp-weights config.yml\n\n\nevaluate\nEvaluates a model’s performance (loss etc) on the train and eval datasets.\n# Basic evaluation\naxolotl evaluate config.yml\n\n\nlm-eval\nRuns LM Evaluation Harness on your model.\n# Basic evaluation\naxolotl lm-eval config.yml\nConfiguration options:\n# List of tasks to evaluate\nlm_eval_tasks:\n  - arc_challenge\n  - hellaswag\nlm_eval_batch_size: # Batch size for evaluation\noutput_dir: # Directory to save evaluation results\nSee LM Eval Harness for more details.\n\n\ndelinearize-llama4\nDelinearizes a Llama 4 linearized model into a regular HuggingFace Llama 4 model. This only works with the non-quantized linearized model.\naxolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir\nThis would be necessary to use with other frameworks. If you have an adapter, merge it with the non-quantized linearized model before delinearizing.\n\n\nquantize\nQuantizes a model using the quantization configuration specified in your YAML file.\naxolotl quantize config.yml\nSee Quantization for more details.",
    "crumbs": [
      "Getting Started",
      "Command Line Interface (CLI)"
    ]
  },
  {
    "objectID": "docs/cli.html#legacy-cli-usage",
    "href": "docs/cli.html#legacy-cli-usage",
    "title": "Command Line Interface (CLI)",
    "section": "Legacy CLI Usage",
    "text": "Legacy CLI Usage\nWhile the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:\n# Preprocess\npython -m axolotl.cli.preprocess config.yml\n\n# Train\naccelerate launch -m axolotl.cli.train config.yml\n\n# Inference\naccelerate launch -m axolotl.cli.inference config.yml \\\n    --lora_model_dir=\"./outputs/lora-out\"\n\n# Gradio interface\naccelerate launch -m axolotl.cli.inference config.yml \\\n    --lora_model_dir=\"./outputs/lora-out\" --gradio\n\n\n\n\n\n\nImportant\n\n\n\nWhen overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., --lora_model_dir).\nNote: This differs from the new Click-based CLI, which uses dash notation (e.g., --lora-model-dir). Keep this in mind if you’re referencing newer documentation or switching between CLI versions.",
    "crumbs": [
      "Getting Started",
      "Command Line Interface (CLI)"
    ]
  },
  {
    "objectID": "docs/cli.html#remote-compute-with-modal-cloud",
    "href": "docs/cli.html#remote-compute-with-modal-cloud",
    "title": "Command Line Interface (CLI)",
    "section": "Remote Compute with Modal Cloud",
    "text": "Remote Compute with Modal Cloud\nAxolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a\ncloud YAML file alongside your regular Axolotl config.\n\nCloud Configuration\nCreate a cloud config YAML with your Modal settings:\n# cloud_config.yml\nprovider: modal\ngpu: a100       # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4\ngpu_count: 1    # Number of GPUs to use\ntimeout: 86400  # Maximum runtime in seconds (24 hours)\nbranch: main    # Git branch to use (optional)\n\nvolumes:        # Persistent storage volumes\n  - name: axolotl-cache\n    mount: /workspace/cache\n  - name: axolotl-data\n    mount: /workspace/data\n  - name: axolotl-artifacts\n    mount: /workspace/artifacts\n\nsecrets:        # Secrets to inject\n  - WANDB_API_KEY\n  - HF_TOKEN\n\n\nRunning on Modal Cloud\nCommands that support the –cloud flag:\n# Preprocess on cloud\naxolotl preprocess config.yml --cloud cloud_config.yml\n\n# Train on cloud\naxolotl train config.yml --cloud cloud_config.yml\n\n# Train without accelerate on cloud\naxolotl train config.yml --cloud cloud_config.yml --no-accelerate\n\n# Run lm-eval on cloud\naxolotl lm-eval config.yml --cloud cloud_config.yml\n\n\nCloud Configuration Options\nprovider:    # compute provider, currently only `modal` is supported\ngpu:         # GPU type to use\ngpu_count:   # Number of GPUs (default: 1)\nmemory:      # RAM in GB (default: 128)\ntimeout:     # Maximum runtime in seconds\ntimeout_preprocess: # Preprocessing timeout\nbranch:      # Git branch to use\ndocker_tag:  # Custom Docker image tag\nvolumes:     # List of persistent storage volumes\n\n# Environment variables to pass. Can be specified in two ways:\n# 1. As a string: Will load the value from the host computer's environment variables\n# 2. As a key-value pair: Will use the specified value directly\n# Example:\n# env:\n#   - CUSTOM_VAR  # Loads from host's $CUSTOM_VAR\n#   - {CUSTOM_VAR: \"value\"}  # Uses \"value\" directly\nenv:\n\n# Secrets to inject. Same input format as `env` but for sensitive data.\nsecrets:\n  # - HF_TOKEN\n  # - WANDB_API_KEY",
    "crumbs": [
      "Getting Started",
      "Command Line Interface (CLI)"
    ]
  },
  {
    "objectID": "docs/multi-node.html",
    "href": "docs/multi-node.html",
    "title": "Multi Node",
    "section": "",
    "text": "The below are three ways to train multi-node in Axolotl.",
    "crumbs": [
      "Deployments",
      "Multi Node"
    ]
  },
  {
    "objectID": "docs/multi-node.html#accelerate",
    "href": "docs/multi-node.html#accelerate",
    "title": "Multi Node",
    "section": "Accelerate",
    "text": "Accelerate\nYou will need to create a configuration for accelerate, either by using accelerate config and follow the instructions or you can use one of the preset below:\n~/.cache/huggingface/accelerate/default_config.yaml\ncompute_environment: LOCAL_MACHINE\ndebug: false\ndistributed_type: FSDP\ndowncast_bf16: 'no'\nmachine_rank: 0 # Set to 0 for the main machine, increment by one for other machines\nmain_process_ip: 10.0.0.4 # Set to main machine's IP\nmain_process_port: 5000\nmain_training_function: main\nmixed_precision: bf16\nnum_machines: 2 # Change to the number of machines\nnum_processes: 4 # That's the total number of GPUs, (for example: if you have 2 machines with 4 GPU, put 8)\nrdzv_backend: static\nsame_network: true\ntpu_env: []\ntpu_use_cluster: false\ntpu_use_sudo: false\nuse_cpu: false\nConfigure your model to use FSDP in the Axolotl yaml. For example:\nfsdp:\n  - full_shard\n  - auto_wrap\nfsdp_config:\n  fsdp_offload_params: true\n  fsdp_state_dict_type: FULL_STATE_DICT\n  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer\nAll you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.",
    "crumbs": [
      "Deployments",
      "Multi Node"
    ]
  },
  {
    "objectID": "docs/multi-node.html#raytrain",
    "href": "docs/multi-node.html#raytrain",
    "title": "Multi Node",
    "section": "Raytrain",
    "text": "Raytrain\nPlease see ray train doc here.",
    "crumbs": [
      "Deployments",
      "Multi Node"
    ]
  },
  {
    "objectID": "docs/multi-node.html#torchrun",
    "href": "docs/multi-node.html#torchrun",
    "title": "Multi Node",
    "section": "Torchrun",
    "text": "Torchrun\nIf you are using Infiniband, we recommend torchrun to utilize the full bandwidth.\nSet the following env (change buffersize/socketname depending on your system):\nexport NCCL_IB_DISABLE=0\nexport NCCL_SOCKET_IFNAME=\"eth0,en,eth,em,bond\"\nexport NCCL_BUFFSIZE=2097152\nRun the following on each node:\ntorchrun --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint \"$head_node_ip:$head_node_port\" -m axolotl.cli.train config.yaml\nPlease make sure to substitute the placeholder variables.\n\nnum_nodes: Number of nodes (containing GPUs)\ngpu_per_node: Number of gpus per node\nhead_node_ip: IP of the head node (make sure other machines can connect to this)\nhead_node_port: Port of the head node (make sure other machines can connect to this. Default 29400)\nrdzv_id: A unique job ID that is used by the job across nodes.\n\n\n\n\n\n\n\nNote\n\n\n\nYou need to call axolotl.cli.train instead of axolotl train as the latter calls accelerate under the hood\n\n\nMore info on the available configs can be found on the Pytorch docs here",
    "crumbs": [
      "Deployments",
      "Multi Node"
    ]
  },
  {
    "objectID": "docs/nccl.html",
    "href": "docs/nccl.html",
    "title": "NCCL",
    "section": "",
    "text": "NVIDIA NCCL is a library to facilitate and optimize multi-GPU communication operations, such as broadcast, all-gather, reduce, all-reduce, etc. Broadly, NCCL configuration is highly environment-specific and is configured via several environment variables. A common NCCL-related problem occurs when a long-running operation times out causing the training process to abort:\nWatchdog caught collective operation timeout: WorkNCCL(SeqNum=42, OpType=ALLGATHER, Timeout(ms)=1800000) ran for 1806948 milliseconds before timing out.\nOften, this timeout will happen after 30 minutes (the default setting) and is accompanied by below-average power consumption with near 100% GPU utilization before the error is raised. Nvidia recommends disabling PCI access control services (ACS) as a possible solution if this is available to you.\nForcing cross-GPU communication via NVLink may help without increasing timeouts. To verify that your configuration is leveraging NVLink run the following command:\nnvidia-smi nvlink --status\nTo force NCCL to use NVLink, simply set this in the environment:\nexport NCCL_P2P_LEVEL=NVL\nIf NVLink is not available in your environment there are other options for NCCL_P2P_LEVEL in the table below:\n\n\n\n\n\n\n\nNCCL_P2P_LEVEL\nDescription\n\n\n\n\nPIX\nP2P data transfers through no more than a single PCIe bridge. Faster data transfer rates vs to paths involving multiple bridges, but slower compared to direct GPU-to-GPU communication.\n\n\nPXB\nP2P data transfers through multiple PCIe bridges but not going through the PCIe Host Bridge; this path involves a complex routing process, potentially incurring a moderate level of latency.\n\n\nPHB\nP2P data transfers occur over the PCIe and through a PCIe Host Bridge, typically involving the CPU, which can facilitate direct memory access but might introduce additional latency compared to more direct paths (ex PIX, NVL)\n\n\n\nTo validate that acceptable data transfer speeds exist for your training job, running NCCL Tests can help pinpoint bottlenecks, for example:\n./build/all_reduce_perf -b 8 -e 128M -f 2 -g 3\nIt can be useful when debugging NCCL communication timeouts to activate additional logging in both PyTorch and NCCL:\nexport NCCL_DEBUG=INFO\nexport NCCL_DEBUG_SUBSYS=ALL\nexport TORCH_DISTRIBUTED_DEBUG=INFO\nexport TORCHELASTIC_ERROR_FILE=/PATH/TO/torcherror.log\nFinally, if you believe your training job needs more time you can increase the timeout past 30 minutes by setting the ddp_timeout value in the Axolotl configuration. See PyTorch init_process_group for documentation on this value.",
    "crumbs": [
      "Troubleshooting",
      "NCCL"
    ]
  },
  {
    "objectID": "docs/api/utils.data.sft.html",
    "href": "docs/api/utils.data.sft.html",
    "title": "utils.data.sft",
    "section": "",
    "text": "utils.data.sft\nutils.data.sft\ndata handling specific to SFT"
  },
  {
    "objectID": "docs/api/core.datasets.transforms.chat_builder.html",
    "href": "docs/api/core.datasets.transforms.chat_builder.html",
    "title": "core.datasets.transforms.chat_builder",
    "section": "",
    "text": "core.datasets.transforms.chat_builder\nThis module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat.\n\n\n\n\n\nName\nDescription\n\n\n\n\nchat_message_transform_builder\nBuilds a transform that takes a row from the dataset and converts it to a Chat\n\n\n\n\n\ncore.datasets.transforms.chat_builder.chat_message_transform_builder(\n    train_on_inputs=False,\n    conversations_field='conversations',\n    message_field_role=['role', 'from'],\n    message_field_content=['value', 'text', 'content'],\n    message_field_training=['train', 'weight'],\n)\nBuilds a transform that takes a row from the dataset and converts it to a Chat\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrain_on_inputs\nbool\nIf True, the transform will train on the inputs. If False, the transform will train on the targets. Defaults to False.\nFalse\n\n\nconversations_field\nstr\nThe field name of the conversations. Defaults to “conversations”.\n'conversations'\n\n\nmessage_field_role\nstr | list[str]\nThe field name of the role. Defaults to “role”.\n['role', 'from']\n\n\nmessage_field_content\nstr | list[str]\nThe field name of the message content. Defaults to “content”.\n['value', 'text', 'content']\n\n\nmessage_field_training\nstr | list[str]\nThe field name of the train/weight. Defaults to “weight”.\n['train', 'weight']\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nCallable\n\nA function that takes a list of conversations and returns a list of messages."
  },
  {
    "objectID": "docs/api/core.datasets.transforms.chat_builder.html#functions",
    "href": "docs/api/core.datasets.transforms.chat_builder.html#functions",
    "title": "core.datasets.transforms.chat_builder",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nchat_message_transform_builder\nBuilds a transform that takes a row from the dataset and converts it to a Chat\n\n\n\n\n\ncore.datasets.transforms.chat_builder.chat_message_transform_builder(\n    train_on_inputs=False,\n    conversations_field='conversations',\n    message_field_role=['role', 'from'],\n    message_field_content=['value', 'text', 'content'],\n    message_field_training=['train', 'weight'],\n)\nBuilds a transform that takes a row from the dataset and converts it to a Chat\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrain_on_inputs\nbool\nIf True, the transform will train on the inputs. If False, the transform will train on the targets. Defaults to False.\nFalse\n\n\nconversations_field\nstr\nThe field name of the conversations. Defaults to “conversations”.\n'conversations'\n\n\nmessage_field_role\nstr | list[str]\nThe field name of the role. Defaults to “role”.\n['role', 'from']\n\n\nmessage_field_content\nstr | list[str]\nThe field name of the message content. Defaults to “content”.\n['value', 'text', 'content']\n\n\nmessage_field_training\nstr | list[str]\nThe field name of the train/weight. Defaults to “weight”.\n['train', 'weight']\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nCallable\n\nA function that takes a list of conversations and returns a list of messages."
  },
  {
    "objectID": "docs/api/evaluate.html",
    "href": "docs/api/evaluate.html",
    "title": "evaluate",
    "section": "",
    "text": "evaluate\nModule for evaluating models.\n\n\n\n\n\nName\nDescription\n\n\n\n\nevaluate\nEvaluate a model on training and validation datasets.\n\n\nevaluate_dataset\nHelper function to evaluate a single dataset.\n\n\n\n\n\nevaluate.evaluate(cfg, dataset_meta)\nEvaluate a model on training and validation datasets.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ndataset_meta\nTrainDatasetMeta\nDataset metadata containing training and evaluation datasets.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nDict[str, float]\nDictionary mapping metric names to their values.\n\n\n\n\n\n\n\nevaluate.evaluate_dataset(trainer, dataset, dataset_type, flash_optimum=False)\nHelper function to evaluate a single dataset.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrainer\nTrainer\nThe trainer instance.\nrequired\n\n\ndataset\nDataset\nDataset to evaluate.\nrequired\n\n\ndataset_type\nstr\nType of dataset (‘train’ or ‘eval’).\nrequired\n\n\nflash_optimum\nbool\nWhether to use flash optimum.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nOptional[Dict[str, float]]\nDictionary of metrics or None if dataset is None."
  },
  {
    "objectID": "docs/api/evaluate.html#functions",
    "href": "docs/api/evaluate.html#functions",
    "title": "evaluate",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nevaluate\nEvaluate a model on training and validation datasets.\n\n\nevaluate_dataset\nHelper function to evaluate a single dataset.\n\n\n\n\n\nevaluate.evaluate(cfg, dataset_meta)\nEvaluate a model on training and validation datasets.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ndataset_meta\nTrainDatasetMeta\nDataset metadata containing training and evaluation datasets.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nDict[str, float]\nDictionary mapping metric names to their values.\n\n\n\n\n\n\n\nevaluate.evaluate_dataset(trainer, dataset, dataset_type, flash_optimum=False)\nHelper function to evaluate a single dataset.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrainer\nTrainer\nThe trainer instance.\nrequired\n\n\ndataset\nDataset\nDataset to evaluate.\nrequired\n\n\ndataset_type\nstr\nType of dataset (‘train’ or ‘eval’).\nrequired\n\n\nflash_optimum\nbool\nWhether to use flash optimum.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nOptional[Dict[str, float]]\nDictionary of metrics or None if dataset is None."
  },
  {
    "objectID": "docs/api/monkeypatch.mixtral.html",
    "href": "docs/api/monkeypatch.mixtral.html",
    "title": "monkeypatch.mixtral",
    "section": "",
    "text": "monkeypatch.mixtral\nmonkeypatch.mixtral\nPatches to support multipack for mixtral"
  },
  {
    "objectID": "docs/api/prompt_strategies.alpaca_w_system.html",
    "href": "docs/api/prompt_strategies.alpaca_w_system.html",
    "title": "prompt_strategies.alpaca_w_system",
    "section": "",
    "text": "prompt_strategies.alpaca_w_system\nPrompt strategies loader for alpaca instruction datasets with system prompts\n\n\n\n\n\nName\nDescription\n\n\n\n\nInstructionWSystemPromptTokenizingStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nOpenOrcaPromptTokenizingStrategy\nTokenizing strategy for OpenOrca datasets\n\n\nOpenOrcaSystemDataPrompter\nAlpaca Style Prompter that uses system prompts from the dataset, with OpenOrca prompts\n\n\nSystemDataPrompter\nAlpaca Style Prompter that uses system prompts from the dataset\n\n\n\n\n\nprompt_strategies.alpaca_w_system.InstructionWSystemPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\nprompt_strategies.alpaca_w_system.OpenOrcaPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for OpenOrca datasets\n\n\n\nprompt_strategies.alpaca_w_system.OpenOrcaSystemDataPrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAlpaca Style Prompter that uses system prompts from the dataset, with OpenOrca prompts\n\n\n\nprompt_strategies.alpaca_w_system.SystemDataPrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAlpaca Style Prompter that uses system prompts from the dataset"
  },
  {
    "objectID": "docs/api/prompt_strategies.alpaca_w_system.html#classes",
    "href": "docs/api/prompt_strategies.alpaca_w_system.html#classes",
    "title": "prompt_strategies.alpaca_w_system",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nInstructionWSystemPromptTokenizingStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nOpenOrcaPromptTokenizingStrategy\nTokenizing strategy for OpenOrca datasets\n\n\nOpenOrcaSystemDataPrompter\nAlpaca Style Prompter that uses system prompts from the dataset, with OpenOrca prompts\n\n\nSystemDataPrompter\nAlpaca Style Prompter that uses system prompts from the dataset\n\n\n\n\n\nprompt_strategies.alpaca_w_system.InstructionWSystemPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\nprompt_strategies.alpaca_w_system.OpenOrcaPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for OpenOrca datasets\n\n\n\nprompt_strategies.alpaca_w_system.OpenOrcaSystemDataPrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAlpaca Style Prompter that uses system prompts from the dataset, with OpenOrca prompts\n\n\n\nprompt_strategies.alpaca_w_system.SystemDataPrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAlpaca Style Prompter that uses system prompts from the dataset"
  },
  {
    "objectID": "docs/api/integrations.spectrum.args.html",
    "href": "docs/api/integrations.spectrum.args.html",
    "title": "integrations.spectrum.args",
    "section": "",
    "text": "integrations.spectrum.args\nModule for handling Spectrum input arguments.\n\n\n\n\n\nName\nDescription\n\n\n\n\nSpectrumArgs\nInput args for Spectrum.\n\n\n\n\n\nintegrations.spectrum.args.SpectrumArgs()\nInput args for Spectrum."
  },
  {
    "objectID": "docs/api/integrations.spectrum.args.html#classes",
    "href": "docs/api/integrations.spectrum.args.html#classes",
    "title": "integrations.spectrum.args",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nSpectrumArgs\nInput args for Spectrum.\n\n\n\n\n\nintegrations.spectrum.args.SpectrumArgs()\nInput args for Spectrum."
  },
  {
    "objectID": "docs/api/core.builders.rl.html",
    "href": "docs/api/core.builders.rl.html",
    "title": "core.builders.rl",
    "section": "",
    "text": "core.builders.rl\nBuilder for RLHF trainers\n\n\n\n\n\nName\nDescription\n\n\n\n\nHFPPOTrainerBuilder\nHF Factory class for PPO Trainer\n\n\nHFRLTrainerBuilder\nTrainer factory class for TRL-based RLHF trainers (e.g. DPO)\n\n\n\n\n\ncore.builders.rl.HFPPOTrainerBuilder(cfg, model, tokenizer, processor=None)\nHF Factory class for PPO Trainer\n\n\n\ncore.builders.rl.HFRLTrainerBuilder(cfg, model, tokenizer, processor=None)\nTrainer factory class for TRL-based RLHF trainers (e.g. DPO)"
  },
  {
    "objectID": "docs/api/core.builders.rl.html#classes",
    "href": "docs/api/core.builders.rl.html#classes",
    "title": "core.builders.rl",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nHFPPOTrainerBuilder\nHF Factory class for PPO Trainer\n\n\nHFRLTrainerBuilder\nTrainer factory class for TRL-based RLHF trainers (e.g. DPO)\n\n\n\n\n\ncore.builders.rl.HFPPOTrainerBuilder(cfg, model, tokenizer, processor=None)\nHF Factory class for PPO Trainer\n\n\n\ncore.builders.rl.HFRLTrainerBuilder(cfg, model, tokenizer, processor=None)\nTrainer factory class for TRL-based RLHF trainers (e.g. DPO)"
  },
  {
    "objectID": "docs/api/common.const.html",
    "href": "docs/api/common.const.html",
    "title": "common.const",
    "section": "",
    "text": "common.const\ncommon.const\nVarious shared constants"
  },
  {
    "objectID": "docs/api/kernels.geglu.html",
    "href": "docs/api/kernels.geglu.html",
    "title": "kernels.geglu",
    "section": "",
    "text": "kernels.geglu\nModule for definition of GEGLU Triton kernels.\nSee “GLU Variants Improve Transformer” (https://arxiv.org/abs/2002.05202).\nCredit to unsloth (https://unsloth.ai/) for inspiration for this implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\ngeglu_backward\nGEGLU backward pass using in-place operations.\n\n\ngeglu_forward\nGEGLU forward pass.\n\n\n\n\n\nkernels.geglu.geglu_backward(grad_output, gate, up)\nGEGLU backward pass using in-place operations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to output, shape [batch, seq_len, hidden_dim].\nrequired\n\n\ngate\ntorch.Tensor\nGate tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple containing: - GEGLU activation output (h) - Gradient with respect to gate (grad_gate) - Gradient with respect to up (grad_up)\n\n\n\n\n\n\nThis function modifies its input tensors in-place to store results.\n\n\n\n\nkernels.geglu.geglu_forward(gate, up)\nGEGLU forward pass.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngate\ntorch.Tensor\nInput gate tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\ntorch.Tensor: Output tensor of shape [batch, seq_len, hidden_dim]."
  },
  {
    "objectID": "docs/api/kernels.geglu.html#functions",
    "href": "docs/api/kernels.geglu.html#functions",
    "title": "kernels.geglu",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ngeglu_backward\nGEGLU backward pass using in-place operations.\n\n\ngeglu_forward\nGEGLU forward pass.\n\n\n\n\n\nkernels.geglu.geglu_backward(grad_output, gate, up)\nGEGLU backward pass using in-place operations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to output, shape [batch, seq_len, hidden_dim].\nrequired\n\n\ngate\ntorch.Tensor\nGate tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple containing: - GEGLU activation output (h) - Gradient with respect to gate (grad_gate) - Gradient with respect to up (grad_up)\n\n\n\n\n\n\nThis function modifies its input tensors in-place to store results.\n\n\n\n\nkernels.geglu.geglu_forward(gate, up)\nGEGLU forward pass.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngate\ntorch.Tensor\nInput gate tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\ntorch.Tensor: Output tensor of shape [batch, seq_len, hidden_dim]."
  },
  {
    "objectID": "docs/api/index.html",
    "href": "docs/api/index.html",
    "title": "API Reference",
    "section": "",
    "text": "Core functionality for training\n\n\n\ntrain\nPrepare and train a model on a dataset. Can also infer from a model or merge lora\n\n\nevaluate\nModule for evaluating models.\n\n\ndatasets\nModule containing Dataset functionality\n\n\nconvert\nModule containing File Reader, File Writer, Json Parser, and Jsonl Serializer classes\n\n\nprompt_tokenizers\nModule containing PromptTokenizingStrategy and Prompter classes\n\n\nlogging_config\nCommon logging module for axolotl\n\n\ncore.builders.base\nBase class for trainer builder\n\n\ncore.builders.causal\nBuilder for causal trainers\n\n\ncore.builders.rl\nBuilder for RLHF trainers\n\n\ncore.training_args\nextra axolotl specific training args\n\n\ncore.chat.messages\ninternal message representations of chat messages\n\n\ncore.chat.format.chatml\nChatML transformation functions for MessageContents\n\n\ncore.chat.format.llama3x\nLlama 3.x chat formatting functions for MessageContents\n\n\ncore.chat.format.shared\nshared functions for format transforms\n\n\ncore.datasets.chat\nchat dataset module\n\n\ncore.datasets.transforms.chat_builder\nThis module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat.\n\n\n\n\n\n\nCommand-line interface\n\n\n\ncli.main\nClick CLI definitions for various axolotl commands.\n\n\ncli.train\nCLI to run training on a model.\n\n\ncli.evaluate\nCLI to run evaluation on a model.\n\n\ncli.args\nModule for axolotl CLI command arguments.\n\n\ncli.checks\nVarious checks for Axolotl CLI.\n\n\ncli.config\nConfiguration loading and processing.\n\n\ncli.inference\nCLI to run inference on a trained model.\n\n\ncli.merge_lora\nCLI to merge a trained LoRA into a base model.\n\n\ncli.merge_sharded_fsdp_weights\nCLI to merge sharded FSDP model checkpoints into a single combined checkpoint.\n\n\ncli.preprocess\nCLI to run preprocessing of a dataset.\n\n\ncli.sweeps\nUtilities for handling sweeps over configs for axolotl train CLI command\n\n\ncli.utils\nUtility methods for axolotl CLI.\n\n\ncli.vllm_serve\nCLI to start the vllm server for online RL\n\n\ncli.cloud.base\nbase class for cloud platforms from cli\n\n\ncli.cloud.modal_\nModal Cloud support from CLI\n\n\ncli.quantize\nCLI to post-training quantize a model using torchao\n\n\n\n\n\n\nTraining implementations\n\n\n\ncore.trainers.base\nModule for customized trainers\n\n\ncore.trainers.trl\nModule for TRL PPO trainer\n\n\ncore.trainers.mamba\nModule for mamba trainer\n\n\ncore.trainers.relora\nModule for ReLoRA trainer\n\n\ncore.trainers.dpo.trainer\nDPO trainer for axolotl\n\n\ncore.trainers.grpo.trainer\nAxolotl GRPO trainers (with and without sequence parallelism handling)\n\n\ncore.trainers.grpo.sampler\nRepeat random sampler (similar to the one implemented in\n\n\ncore.trainers.utils\nUtils for Axolotl trainers\n\n\n\n\n\n\nFunctionality for loading and patching models, tokenizers, etc.\n\n\n\nloaders.model\nModel loader class implementation for loading, configuring, and patching various\n\n\nloaders.tokenizer\nTokenizer loading functionality and associated utils\n\n\nloaders.processor\nProcessor loading functionality for multi-modal models\n\n\nloaders.adapter\nAdapter loading functionality, including LoRA / QLoRA and associated utils\n\n\nloaders.patch_manager\nPatch manager class implementation to complement axolotl.loaders.ModelLoader.\n\n\nloaders.constants\nShared constants for axolotl.loaders module\n\n\n\n\n\n\nMixin classes for augmenting trainers\n\n\n\ncore.trainers.mixins.optimizer\nModule for Axolotl trainer optimizer mixin\n\n\ncore.trainers.mixins.rng_state_loader\nTemporary fix/override for bug in resume from checkpoint\n\n\ncore.trainers.mixins.scheduler\nModule for Axolotl trainer scheduler mixin\n\n\n\n\n\n\nContext managers for altering trainer behaviors\n\n\n\nutils.ctx_managers.sequence_parallel\nModule for Axolotl trainer sequence parallelism manager and utilities\n\n\n\n\n\n\nPrompt formatting strategies\n\n\n\nprompt_strategies.base\nmodule for base dataset transform strategies\n\n\nprompt_strategies.chat_template\nHF Chat Templates prompt strategy\n\n\nprompt_strategies.alpaca_chat\nModule for Alpaca prompt strategy classes\n\n\nprompt_strategies.alpaca_instruct\nModule loading the AlpacaInstructPromptTokenizingStrategy class\n\n\nprompt_strategies.alpaca_w_system\nPrompt strategies loader for alpaca instruction datasets with system prompts\n\n\nprompt_strategies.user_defined\nUser Defined prompts with configuration from the YML config\n\n\nprompt_strategies.llama2_chat\nPrompt Strategy for finetuning Llama2 chat models\n\n\nprompt_strategies.completion\nBasic completion text\n\n\nprompt_strategies.input_output\nModule for plain input/output prompt pairs\n\n\nprompt_strategies.stepwise_supervised\nModule for stepwise datasets, typically including a prompt and reasoning traces,\n\n\nprompt_strategies.metharme\nModule containing the MetharmenPromptTokenizingStrategy and MetharmePrompter class\n\n\nprompt_strategies.orcamini\nPrompt Strategy for finetuning Orca Mini (v2) models\n\n\nprompt_strategies.pygmalion\nModule containing the PygmalionPromptTokenizingStrategy and PygmalionPrompter class\n\n\nprompt_strategies.messages.chat\nChat dataset wrapping strategy for new internal messages representations\n\n\nprompt_strategies.dpo.chat_template\nDPO prompt strategies for using tokenizer chat templates.\n\n\nprompt_strategies.dpo.llama3\nDPO strategies for llama-3 chat template\n\n\nprompt_strategies.dpo.chatml\nDPO strategies for chatml\n\n\nprompt_strategies.dpo.zephyr\nDPO strategies for zephyr\n\n\nprompt_strategies.dpo.user_defined\nUser-defined DPO strategies\n\n\nprompt_strategies.dpo.passthrough\nDPO prompt strategies passthrough/zero-processing strategy\n\n\nprompt_strategies.kto.llama3\nKTO strategies for llama-3 chat template\n\n\nprompt_strategies.kto.chatml\nKTO strategies for chatml\n\n\nprompt_strategies.kto.user_defined\nUser-defined KTO strategies\n\n\nprompt_strategies.orpo.chat_template\nchatml prompt tokenization strategy for ORPO\n\n\nprompt_strategies.bradley_terry.llama3\nchatml transforms for datasets with system, input, chosen, rejected to match llama3 chat template\n\n\n\n\n\n\nLow-level performance optimizations\n\n\n\nkernels.lora\nModule for definition of Low-Rank Adaptation (LoRA) Triton kernels.\n\n\nkernels.geglu\nModule for definition of GEGLU Triton kernels.\n\n\nkernels.swiglu\nModule for definition of SwiGLU Triton kernels.\n\n\nkernels.quantize\nDequantization utilities for bitsandbytes integration.\n\n\nkernels.utils\nUtilities for axolotl.kernels submodules.\n\n\n\n\n\n\nRuntime patches for model optimizations\n\n\n\nmonkeypatch.llama_attn_hijack_flash\nFlash attention monkey patch for llama model\n\n\nmonkeypatch.llama_attn_hijack_xformers\nDirectly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments\n\n\nmonkeypatch.mistral_attn_hijack_flash\nFlash attention monkey patch for mistral model\n\n\nmonkeypatch.multipack\nmultipack patching for v2 of sample packing\n\n\nmonkeypatch.relora\nImplements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune.\n\n\nmonkeypatch.llama_expand_mask\nexpands the binary attention mask per 3.2.2 of https://arxiv.org/pdf/2107.02027.pdf\n\n\nmonkeypatch.lora_kernels\nModule for patching custom LoRA Triton kernels and torch.autograd functions.\n\n\nmonkeypatch.utils\nShared utils for the monkeypatches\n\n\nmonkeypatch.btlm_attn_hijack_flash\nFlash attention monkey patch for cerebras btlm model\n\n\nmonkeypatch.llama_patch_multipack\nPatched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention\n\n\nmonkeypatch.stablelm_attn_hijack_flash\nPyTorch StableLM Epoch model.\n\n\nmonkeypatch.trainer_fsdp_optim\nfix for FSDP optimizer save in trainer w 4.47.0\n\n\nmonkeypatch.transformers_fa_utils\nsee https://github.com/huggingface/transformers/pull/35834\n\n\nmonkeypatch.unsloth_\nmodule for patching with unsloth optimizations\n\n\nmonkeypatch.attention.mllama\nMonkeypatch for Vision Llama for FA2 support\n\n\nmonkeypatch.data.batch_dataset_fetcher\nmonkey patches for the dataset fetcher to handle batches of packed indexes\n\n\nmonkeypatch.mixtral\nPatches to support multipack for mixtral\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu\nCPU offloaded checkpointing\n\n\nmonkeypatch.gradient_checkpointing.offload_disk\nDISCO - DIsk-based Storage and Checkpointing with Optimized prefetching\n\n\n\n\n\n\nUtility functions\n\n\n\nutils.tokenization\nModule for tokenization utilities\n\n\nutils.chat_templates\nThis module provides functionality for selecting chat templates based on user choices.\n\n\nutils.lora\nmodule to get the state dict of a merged lora model\n\n\nutils.model_shard_quant\nmodule to handle loading model on cpu/meta device for FSDP\n\n\nutils.bench\nBenchmarking and measurement utilities\n\n\nutils.freeze\nmodule to freeze/unfreeze parameters by name\n\n\nutils.trainer\nModule containing the Trainer class and related functions\n\n\nutils.schedulers\nModule for custom LRScheduler class\n\n\nutils.distributed\nutility helpers for distributed checks\n\n\nutils.dict\nModule containing the DictDefault class\n\n\nutils.optimizers.adopt\nCopied from https://github.com/iShohei220/adopt\n\n\nutils.data.pretraining\ndata handling specific to pretraining\n\n\nutils.data.sft\ndata handling specific to SFT\n\n\nutils.quantization\nUtilities for quantization including QAT and PTQ using torchao.\n\n\n\n\n\n\nPydantic data models for Axolotl config\n\n\n\nutils.schemas.config\nModule with Pydantic models for configuration.\n\n\nutils.schemas.model\nPydantic models for model input / output, etc. configuration\n\n\nutils.schemas.training\nPydantic models for training hyperparameters\n\n\nutils.schemas.datasets\nPydantic models for datasets-related configuration\n\n\nutils.schemas.peft\nPydantic models for PEFT-related configuration\n\n\nutils.schemas.trl\nPydantic models for TRL trainer configuration\n\n\nutils.schemas.multimodal\nPydantic models for multimodal-related configuration\n\n\nutils.schemas.integrations\nPydantic models for Axolotl integrations\n\n\nutils.schemas.enums\nEnums for Axolotl input config\n\n\nutils.schemas.utils\nUtilities for Axolotl Pydantic models\n\n\n\n\n\n\nThird-party integrations and extensions\n\n\n\nintegrations.base\nBase class for all plugins.\n\n\nintegrations.cut_cross_entropy.args\nModule for handling Cut Cross Entropy input arguments.\n\n\nintegrations.grokfast.optimizer\n\n\n\nintegrations.kd.trainer\nKD trainer\n\n\nintegrations.liger.args\nModule for handling LIGER input arguments.\n\n\nintegrations.lm_eval.args\nModule for handling lm eval harness input arguments.\n\n\nintegrations.spectrum.args\nModule for handling Spectrum input arguments.\n\n\n\n\n\n\nCommon utilities and shared functionality\n\n\n\ncommon.architectures\nCommon architecture specific constants\n\n\ncommon.const\nVarious shared constants\n\n\ncommon.datasets\nDataset loading utilities.\n\n\n\n\n\n\nCustom model implementations\n\n\n\nmodels.mamba.modeling_mamba\n\n\n\n\n\n\n\nData processing utilities\n\n\n\nutils.collators.core\nbasic shared collator constants\n\n\nutils.collators.batching\nData collators for axolotl to pad labels and position_ids for packed sequences\n\n\nutils.collators.mamba\ncollators for Mamba\n\n\nutils.collators.mm_chat\nCollators for multi-modal chat messages and packing\n\n\nutils.samplers.multipack\nMultipack Batch Sampler - An efficient batch sampler for packing variable-length sequences\n\n\n\n\n\n\nTraining callbacks\n\n\n\nutils.callbacks.perplexity\ncallback to calculate perplexity as an evaluation metric.\n\n\nutils.callbacks.profiler\nHF Trainer callback for creating pytorch profiling snapshots\n\n\nutils.callbacks.lisa\nmodule for LISA\n\n\nutils.callbacks.mlflow_\nMLFlow module for trainer callbacks\n\n\nutils.callbacks.comet_\nComet module for trainer callbacks\n\n\nutils.callbacks.qat\nQAT Callback for HF Causal Trainer"
  },
  {
    "objectID": "docs/api/index.html#core",
    "href": "docs/api/index.html#core",
    "title": "API Reference",
    "section": "",
    "text": "Core functionality for training\n\n\n\ntrain\nPrepare and train a model on a dataset. Can also infer from a model or merge lora\n\n\nevaluate\nModule for evaluating models.\n\n\ndatasets\nModule containing Dataset functionality\n\n\nconvert\nModule containing File Reader, File Writer, Json Parser, and Jsonl Serializer classes\n\n\nprompt_tokenizers\nModule containing PromptTokenizingStrategy and Prompter classes\n\n\nlogging_config\nCommon logging module for axolotl\n\n\ncore.builders.base\nBase class for trainer builder\n\n\ncore.builders.causal\nBuilder for causal trainers\n\n\ncore.builders.rl\nBuilder for RLHF trainers\n\n\ncore.training_args\nextra axolotl specific training args\n\n\ncore.chat.messages\ninternal message representations of chat messages\n\n\ncore.chat.format.chatml\nChatML transformation functions for MessageContents\n\n\ncore.chat.format.llama3x\nLlama 3.x chat formatting functions for MessageContents\n\n\ncore.chat.format.shared\nshared functions for format transforms\n\n\ncore.datasets.chat\nchat dataset module\n\n\ncore.datasets.transforms.chat_builder\nThis module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat."
  },
  {
    "objectID": "docs/api/index.html#cli",
    "href": "docs/api/index.html#cli",
    "title": "API Reference",
    "section": "",
    "text": "Command-line interface\n\n\n\ncli.main\nClick CLI definitions for various axolotl commands.\n\n\ncli.train\nCLI to run training on a model.\n\n\ncli.evaluate\nCLI to run evaluation on a model.\n\n\ncli.args\nModule for axolotl CLI command arguments.\n\n\ncli.checks\nVarious checks for Axolotl CLI.\n\n\ncli.config\nConfiguration loading and processing.\n\n\ncli.inference\nCLI to run inference on a trained model.\n\n\ncli.merge_lora\nCLI to merge a trained LoRA into a base model.\n\n\ncli.merge_sharded_fsdp_weights\nCLI to merge sharded FSDP model checkpoints into a single combined checkpoint.\n\n\ncli.preprocess\nCLI to run preprocessing of a dataset.\n\n\ncli.sweeps\nUtilities for handling sweeps over configs for axolotl train CLI command\n\n\ncli.utils\nUtility methods for axolotl CLI.\n\n\ncli.vllm_serve\nCLI to start the vllm server for online RL\n\n\ncli.cloud.base\nbase class for cloud platforms from cli\n\n\ncli.cloud.modal_\nModal Cloud support from CLI\n\n\ncli.quantize\nCLI to post-training quantize a model using torchao"
  },
  {
    "objectID": "docs/api/index.html#trainers",
    "href": "docs/api/index.html#trainers",
    "title": "API Reference",
    "section": "",
    "text": "Training implementations\n\n\n\ncore.trainers.base\nModule for customized trainers\n\n\ncore.trainers.trl\nModule for TRL PPO trainer\n\n\ncore.trainers.mamba\nModule for mamba trainer\n\n\ncore.trainers.relora\nModule for ReLoRA trainer\n\n\ncore.trainers.dpo.trainer\nDPO trainer for axolotl\n\n\ncore.trainers.grpo.trainer\nAxolotl GRPO trainers (with and without sequence parallelism handling)\n\n\ncore.trainers.grpo.sampler\nRepeat random sampler (similar to the one implemented in\n\n\ncore.trainers.utils\nUtils for Axolotl trainers"
  },
  {
    "objectID": "docs/api/index.html#model-loading",
    "href": "docs/api/index.html#model-loading",
    "title": "API Reference",
    "section": "",
    "text": "Functionality for loading and patching models, tokenizers, etc.\n\n\n\nloaders.model\nModel loader class implementation for loading, configuring, and patching various\n\n\nloaders.tokenizer\nTokenizer loading functionality and associated utils\n\n\nloaders.processor\nProcessor loading functionality for multi-modal models\n\n\nloaders.adapter\nAdapter loading functionality, including LoRA / QLoRA and associated utils\n\n\nloaders.patch_manager\nPatch manager class implementation to complement axolotl.loaders.ModelLoader.\n\n\nloaders.constants\nShared constants for axolotl.loaders module"
  },
  {
    "objectID": "docs/api/index.html#mixins",
    "href": "docs/api/index.html#mixins",
    "title": "API Reference",
    "section": "",
    "text": "Mixin classes for augmenting trainers\n\n\n\ncore.trainers.mixins.optimizer\nModule for Axolotl trainer optimizer mixin\n\n\ncore.trainers.mixins.rng_state_loader\nTemporary fix/override for bug in resume from checkpoint\n\n\ncore.trainers.mixins.scheduler\nModule for Axolotl trainer scheduler mixin"
  },
  {
    "objectID": "docs/api/index.html#context-managers",
    "href": "docs/api/index.html#context-managers",
    "title": "API Reference",
    "section": "",
    "text": "Context managers for altering trainer behaviors\n\n\n\nutils.ctx_managers.sequence_parallel\nModule for Axolotl trainer sequence parallelism manager and utilities"
  },
  {
    "objectID": "docs/api/index.html#prompt-strategies",
    "href": "docs/api/index.html#prompt-strategies",
    "title": "API Reference",
    "section": "",
    "text": "Prompt formatting strategies\n\n\n\nprompt_strategies.base\nmodule for base dataset transform strategies\n\n\nprompt_strategies.chat_template\nHF Chat Templates prompt strategy\n\n\nprompt_strategies.alpaca_chat\nModule for Alpaca prompt strategy classes\n\n\nprompt_strategies.alpaca_instruct\nModule loading the AlpacaInstructPromptTokenizingStrategy class\n\n\nprompt_strategies.alpaca_w_system\nPrompt strategies loader for alpaca instruction datasets with system prompts\n\n\nprompt_strategies.user_defined\nUser Defined prompts with configuration from the YML config\n\n\nprompt_strategies.llama2_chat\nPrompt Strategy for finetuning Llama2 chat models\n\n\nprompt_strategies.completion\nBasic completion text\n\n\nprompt_strategies.input_output\nModule for plain input/output prompt pairs\n\n\nprompt_strategies.stepwise_supervised\nModule for stepwise datasets, typically including a prompt and reasoning traces,\n\n\nprompt_strategies.metharme\nModule containing the MetharmenPromptTokenizingStrategy and MetharmePrompter class\n\n\nprompt_strategies.orcamini\nPrompt Strategy for finetuning Orca Mini (v2) models\n\n\nprompt_strategies.pygmalion\nModule containing the PygmalionPromptTokenizingStrategy and PygmalionPrompter class\n\n\nprompt_strategies.messages.chat\nChat dataset wrapping strategy for new internal messages representations\n\n\nprompt_strategies.dpo.chat_template\nDPO prompt strategies for using tokenizer chat templates.\n\n\nprompt_strategies.dpo.llama3\nDPO strategies for llama-3 chat template\n\n\nprompt_strategies.dpo.chatml\nDPO strategies for chatml\n\n\nprompt_strategies.dpo.zephyr\nDPO strategies for zephyr\n\n\nprompt_strategies.dpo.user_defined\nUser-defined DPO strategies\n\n\nprompt_strategies.dpo.passthrough\nDPO prompt strategies passthrough/zero-processing strategy\n\n\nprompt_strategies.kto.llama3\nKTO strategies for llama-3 chat template\n\n\nprompt_strategies.kto.chatml\nKTO strategies for chatml\n\n\nprompt_strategies.kto.user_defined\nUser-defined KTO strategies\n\n\nprompt_strategies.orpo.chat_template\nchatml prompt tokenization strategy for ORPO\n\n\nprompt_strategies.bradley_terry.llama3\nchatml transforms for datasets with system, input, chosen, rejected to match llama3 chat template"
  },
  {
    "objectID": "docs/api/index.html#kernels",
    "href": "docs/api/index.html#kernels",
    "title": "API Reference",
    "section": "",
    "text": "Low-level performance optimizations\n\n\n\nkernels.lora\nModule for definition of Low-Rank Adaptation (LoRA) Triton kernels.\n\n\nkernels.geglu\nModule for definition of GEGLU Triton kernels.\n\n\nkernels.swiglu\nModule for definition of SwiGLU Triton kernels.\n\n\nkernels.quantize\nDequantization utilities for bitsandbytes integration.\n\n\nkernels.utils\nUtilities for axolotl.kernels submodules."
  },
  {
    "objectID": "docs/api/index.html#monkey-patches",
    "href": "docs/api/index.html#monkey-patches",
    "title": "API Reference",
    "section": "",
    "text": "Runtime patches for model optimizations\n\n\n\nmonkeypatch.llama_attn_hijack_flash\nFlash attention monkey patch for llama model\n\n\nmonkeypatch.llama_attn_hijack_xformers\nDirectly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments\n\n\nmonkeypatch.mistral_attn_hijack_flash\nFlash attention monkey patch for mistral model\n\n\nmonkeypatch.multipack\nmultipack patching for v2 of sample packing\n\n\nmonkeypatch.relora\nImplements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune.\n\n\nmonkeypatch.llama_expand_mask\nexpands the binary attention mask per 3.2.2 of https://arxiv.org/pdf/2107.02027.pdf\n\n\nmonkeypatch.lora_kernels\nModule for patching custom LoRA Triton kernels and torch.autograd functions.\n\n\nmonkeypatch.utils\nShared utils for the monkeypatches\n\n\nmonkeypatch.btlm_attn_hijack_flash\nFlash attention monkey patch for cerebras btlm model\n\n\nmonkeypatch.llama_patch_multipack\nPatched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention\n\n\nmonkeypatch.stablelm_attn_hijack_flash\nPyTorch StableLM Epoch model.\n\n\nmonkeypatch.trainer_fsdp_optim\nfix for FSDP optimizer save in trainer w 4.47.0\n\n\nmonkeypatch.transformers_fa_utils\nsee https://github.com/huggingface/transformers/pull/35834\n\n\nmonkeypatch.unsloth_\nmodule for patching with unsloth optimizations\n\n\nmonkeypatch.attention.mllama\nMonkeypatch for Vision Llama for FA2 support\n\n\nmonkeypatch.data.batch_dataset_fetcher\nmonkey patches for the dataset fetcher to handle batches of packed indexes\n\n\nmonkeypatch.mixtral\nPatches to support multipack for mixtral\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu\nCPU offloaded checkpointing\n\n\nmonkeypatch.gradient_checkpointing.offload_disk\nDISCO - DIsk-based Storage and Checkpointing with Optimized prefetching"
  },
  {
    "objectID": "docs/api/index.html#utils",
    "href": "docs/api/index.html#utils",
    "title": "API Reference",
    "section": "",
    "text": "Utility functions\n\n\n\nutils.tokenization\nModule for tokenization utilities\n\n\nutils.chat_templates\nThis module provides functionality for selecting chat templates based on user choices.\n\n\nutils.lora\nmodule to get the state dict of a merged lora model\n\n\nutils.model_shard_quant\nmodule to handle loading model on cpu/meta device for FSDP\n\n\nutils.bench\nBenchmarking and measurement utilities\n\n\nutils.freeze\nmodule to freeze/unfreeze parameters by name\n\n\nutils.trainer\nModule containing the Trainer class and related functions\n\n\nutils.schedulers\nModule for custom LRScheduler class\n\n\nutils.distributed\nutility helpers for distributed checks\n\n\nutils.dict\nModule containing the DictDefault class\n\n\nutils.optimizers.adopt\nCopied from https://github.com/iShohei220/adopt\n\n\nutils.data.pretraining\ndata handling specific to pretraining\n\n\nutils.data.sft\ndata handling specific to SFT\n\n\nutils.quantization\nUtilities for quantization including QAT and PTQ using torchao."
  },
  {
    "objectID": "docs/api/index.html#schemas",
    "href": "docs/api/index.html#schemas",
    "title": "API Reference",
    "section": "",
    "text": "Pydantic data models for Axolotl config\n\n\n\nutils.schemas.config\nModule with Pydantic models for configuration.\n\n\nutils.schemas.model\nPydantic models for model input / output, etc. configuration\n\n\nutils.schemas.training\nPydantic models for training hyperparameters\n\n\nutils.schemas.datasets\nPydantic models for datasets-related configuration\n\n\nutils.schemas.peft\nPydantic models for PEFT-related configuration\n\n\nutils.schemas.trl\nPydantic models for TRL trainer configuration\n\n\nutils.schemas.multimodal\nPydantic models for multimodal-related configuration\n\n\nutils.schemas.integrations\nPydantic models for Axolotl integrations\n\n\nutils.schemas.enums\nEnums for Axolotl input config\n\n\nutils.schemas.utils\nUtilities for Axolotl Pydantic models"
  },
  {
    "objectID": "docs/api/index.html#integrations",
    "href": "docs/api/index.html#integrations",
    "title": "API Reference",
    "section": "",
    "text": "Third-party integrations and extensions\n\n\n\nintegrations.base\nBase class for all plugins.\n\n\nintegrations.cut_cross_entropy.args\nModule for handling Cut Cross Entropy input arguments.\n\n\nintegrations.grokfast.optimizer\n\n\n\nintegrations.kd.trainer\nKD trainer\n\n\nintegrations.liger.args\nModule for handling LIGER input arguments.\n\n\nintegrations.lm_eval.args\nModule for handling lm eval harness input arguments.\n\n\nintegrations.spectrum.args\nModule for handling Spectrum input arguments."
  },
  {
    "objectID": "docs/api/index.html#common",
    "href": "docs/api/index.html#common",
    "title": "API Reference",
    "section": "",
    "text": "Common utilities and shared functionality\n\n\n\ncommon.architectures\nCommon architecture specific constants\n\n\ncommon.const\nVarious shared constants\n\n\ncommon.datasets\nDataset loading utilities."
  },
  {
    "objectID": "docs/api/index.html#models",
    "href": "docs/api/index.html#models",
    "title": "API Reference",
    "section": "",
    "text": "Custom model implementations\n\n\n\nmodels.mamba.modeling_mamba"
  },
  {
    "objectID": "docs/api/index.html#data-processing",
    "href": "docs/api/index.html#data-processing",
    "title": "API Reference",
    "section": "",
    "text": "Data processing utilities\n\n\n\nutils.collators.core\nbasic shared collator constants\n\n\nutils.collators.batching\nData collators for axolotl to pad labels and position_ids for packed sequences\n\n\nutils.collators.mamba\ncollators for Mamba\n\n\nutils.collators.mm_chat\nCollators for multi-modal chat messages and packing\n\n\nutils.samplers.multipack\nMultipack Batch Sampler - An efficient batch sampler for packing variable-length sequences"
  },
  {
    "objectID": "docs/api/index.html#callbacks",
    "href": "docs/api/index.html#callbacks",
    "title": "API Reference",
    "section": "",
    "text": "Training callbacks\n\n\n\nutils.callbacks.perplexity\ncallback to calculate perplexity as an evaluation metric.\n\n\nutils.callbacks.profiler\nHF Trainer callback for creating pytorch profiling snapshots\n\n\nutils.callbacks.lisa\nmodule for LISA\n\n\nutils.callbacks.mlflow_\nMLFlow module for trainer callbacks\n\n\nutils.callbacks.comet_\nComet module for trainer callbacks\n\n\nutils.callbacks.qat\nQAT Callback for HF Causal Trainer"
  },
  {
    "objectID": "docs/api/integrations.cut_cross_entropy.args.html",
    "href": "docs/api/integrations.cut_cross_entropy.args.html",
    "title": "integrations.cut_cross_entropy.args",
    "section": "",
    "text": "integrations.cut_cross_entropy.args\nModule for handling Cut Cross Entropy input arguments.\n\n\n\n\n\nName\nDescription\n\n\n\n\nCutCrossEntropyArgs\nInput args for Cut Cross Entropy.\n\n\n\n\n\nintegrations.cut_cross_entropy.args.CutCrossEntropyArgs()\nInput args for Cut Cross Entropy."
  },
  {
    "objectID": "docs/api/integrations.cut_cross_entropy.args.html#classes",
    "href": "docs/api/integrations.cut_cross_entropy.args.html#classes",
    "title": "integrations.cut_cross_entropy.args",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nCutCrossEntropyArgs\nInput args for Cut Cross Entropy.\n\n\n\n\n\nintegrations.cut_cross_entropy.args.CutCrossEntropyArgs()\nInput args for Cut Cross Entropy."
  },
  {
    "objectID": "docs/api/prompt_strategies.user_defined.html",
    "href": "docs/api/prompt_strategies.user_defined.html",
    "title": "prompt_strategies.user_defined",
    "section": "",
    "text": "prompt_strategies.user_defined\nUser Defined prompts with configuration from the YML config\n\n\n\n\n\nName\nDescription\n\n\n\n\nUserDefinedDatasetConfig\ndataclass configuration representing a userdefined dataset type\n\n\nUserDefinedPromptTokenizationStrategy\nPrompt Tokenization Strategy for user defined prompts\n\n\n\n\n\nprompt_strategies.user_defined.UserDefinedDatasetConfig(\n    system_prompt='',\n    field_system='system',\n    field_instruction='instruction',\n    field_input='input',\n    field_output='output',\n    format='{instruction} {input} ',\n    no_input_format='{instruction} ',\n    system_format='{system}',\n)\ndataclass configuration representing a userdefined dataset type\n\n\n\nprompt_strategies.user_defined.UserDefinedPromptTokenizationStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nPrompt Tokenization Strategy for user defined prompts"
  },
  {
    "objectID": "docs/api/prompt_strategies.user_defined.html#classes",
    "href": "docs/api/prompt_strategies.user_defined.html#classes",
    "title": "prompt_strategies.user_defined",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nUserDefinedDatasetConfig\ndataclass configuration representing a userdefined dataset type\n\n\nUserDefinedPromptTokenizationStrategy\nPrompt Tokenization Strategy for user defined prompts\n\n\n\n\n\nprompt_strategies.user_defined.UserDefinedDatasetConfig(\n    system_prompt='',\n    field_system='system',\n    field_instruction='instruction',\n    field_input='input',\n    field_output='output',\n    format='{instruction} {input} ',\n    no_input_format='{instruction} ',\n    system_format='{system}',\n)\ndataclass configuration representing a userdefined dataset type\n\n\n\nprompt_strategies.user_defined.UserDefinedPromptTokenizationStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nPrompt Tokenization Strategy for user defined prompts"
  },
  {
    "objectID": "docs/api/core.trainers.mixins.scheduler.html",
    "href": "docs/api/core.trainers.mixins.scheduler.html",
    "title": "core.trainers.mixins.scheduler",
    "section": "",
    "text": "core.trainers.mixins.scheduler\nModule for Axolotl trainer scheduler mixin\n\n\n\n\n\nName\nDescription\n\n\n\n\nSchedulerMixin\nMixin class for scheduler setup in CausalTrainer.\n\n\n\n\n\ncore.trainers.mixins.scheduler.SchedulerMixin()\nMixin class for scheduler setup in CausalTrainer.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncreate_scheduler\nSet up the scheduler. The optimizer of the trainer must have been set up either before this method is called or\n\n\n\n\n\ncore.trainers.mixins.scheduler.SchedulerMixin.create_scheduler(\n    num_training_steps,\n    optimizer=None,\n)\nSet up the scheduler. The optimizer of the trainer must have been set up either before this method is called or\npassed as an argument.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nnum_training_steps\nint\nThe number of training steps to do.\nrequired\n\n\noptimizer\ntorch.optim.Optimizer\nThe training optimizer\nNone"
  },
  {
    "objectID": "docs/api/core.trainers.mixins.scheduler.html#classes",
    "href": "docs/api/core.trainers.mixins.scheduler.html#classes",
    "title": "core.trainers.mixins.scheduler",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nSchedulerMixin\nMixin class for scheduler setup in CausalTrainer.\n\n\n\n\n\ncore.trainers.mixins.scheduler.SchedulerMixin()\nMixin class for scheduler setup in CausalTrainer.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncreate_scheduler\nSet up the scheduler. The optimizer of the trainer must have been set up either before this method is called or\n\n\n\n\n\ncore.trainers.mixins.scheduler.SchedulerMixin.create_scheduler(\n    num_training_steps,\n    optimizer=None,\n)\nSet up the scheduler. The optimizer of the trainer must have been set up either before this method is called or\npassed as an argument.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nnum_training_steps\nint\nThe number of training steps to do.\nrequired\n\n\noptimizer\ntorch.optim.Optimizer\nThe training optimizer\nNone"
  },
  {
    "objectID": "docs/api/core.datasets.chat.html",
    "href": "docs/api/core.datasets.chat.html",
    "title": "core.datasets.chat",
    "section": "",
    "text": "core.datasets.chat\nchat dataset module\n\n\n\n\n\nName\nDescription\n\n\n\n\nTokenizedChatDataset\nTokenized chat dataset\n\n\n\n\n\ncore.datasets.chat.TokenizedChatDataset(\n    data,\n    model_transform,\n    *args,\n    message_transform=None,\n    formatter=None,\n    process_count=None,\n    keep_in_memory=False,\n    **kwargs,\n)\nTokenized chat dataset"
  },
  {
    "objectID": "docs/api/core.datasets.chat.html#classes",
    "href": "docs/api/core.datasets.chat.html#classes",
    "title": "core.datasets.chat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nTokenizedChatDataset\nTokenized chat dataset\n\n\n\n\n\ncore.datasets.chat.TokenizedChatDataset(\n    data,\n    model_transform,\n    *args,\n    message_transform=None,\n    formatter=None,\n    process_count=None,\n    keep_in_memory=False,\n    **kwargs,\n)\nTokenized chat dataset"
  },
  {
    "objectID": "docs/api/train.html",
    "href": "docs/api/train.html",
    "title": "train",
    "section": "",
    "text": "train\nPrepare and train a model on a dataset. Can also infer from a model or merge lora\n\n\n\n\n\nName\nDescription\n\n\n\n\ncreate_model_card\nCreate a model card for the trained model if needed.\n\n\ndetermine_resume_checkpoint\nDetermine the checkpoint to resume from based on configuration.\n\n\nexecute_training\nExecute the training process with appropriate SDP kernel configurations.\n\n\nhandle_untrained_tokens_fix\nApply fixes for untrained tokens if configured.\n\n\nsave_initial_configs\nSave initial configurations before training.\n\n\nsave_trained_model\nSave the trained model according to configuration and training setup.\n\n\nsetup_model_and_tokenizer\nLoad the tokenizer, processor (for multimodal models), and model based on configuration.\n\n\nsetup_model_and_trainer\nLoad model, tokenizer, trainer, etc. Helper function to encapsulate the full\n\n\nsetup_model_card\nSet up the Axolotl badge and add the Axolotl config to the model card if available.\n\n\nsetup_reference_model\nSet up the reference model for RL training if needed.\n\n\nsetup_signal_handler\nSet up signal handler for graceful termination.\n\n\ntrain\nTrain a model on the given dataset.\n\n\n\n\n\ntrain.create_model_card(cfg, trainer)\nCreate a model card for the trained model if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object with model card creation capabilities.\nrequired\n\n\n\n\n\n\n\ntrain.determine_resume_checkpoint(cfg)\nDetermine the checkpoint to resume from based on configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr | None\nPath to the checkpoint to resume from, or None if not resuming.\n\n\n\n\n\n\n\ntrain.execute_training(cfg, trainer, resume_from_checkpoint)\nExecute the training process with appropriate SDP kernel configurations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntrainer\nAny\nThe configured trainer object.\nrequired\n\n\nresume_from_checkpoint\nstr | None\nPath to checkpoint to resume from, if applicable.\nrequired\n\n\n\n\n\n\n\ntrain.handle_untrained_tokens_fix(\n    cfg,\n    model,\n    tokenizer,\n    train_dataset,\n    safe_serialization,\n)\nApply fixes for untrained tokens if configured.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\nmodel\nPreTrainedModel\nThe model to apply fixes to.\nrequired\n\n\ntokenizer\nPreTrainedTokenizer\nThe tokenizer for token identification.\nrequired\n\n\ntrain_dataset\nDataset\nThe training dataset to use.\nrequired\n\n\nsafe_serialization\nbool\nWhether to use safe serialization when saving.\nrequired\n\n\n\n\n\n\n\ntrain.save_initial_configs(cfg, tokenizer, model, peft_config, processor)\nSave initial configurations before training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntokenizer\nPreTrainedTokenizer\nThe tokenizer to save.\nrequired\n\n\nmodel\nPreTrainedModel\nThe model to save configuration for.\nrequired\n\n\npeft_config\nPeftConfig | None\nThe PEFT configuration to save if applicable.\nrequired\n\n\n\n\n\n\n\ntrain.save_trained_model(cfg, trainer, model, safe_serialization)\nSave the trained model according to configuration and training setup.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntrainer\nAny\nThe trainer object.\nrequired\n\n\nmodel\nPreTrainedModel\nThe trained model to save.\nrequired\n\n\nsafe_serialization\nbool\nWhether to use safe serialization.\nrequired\n\n\n\n\n\n\n\ntrain.setup_model_and_tokenizer(cfg)\nLoad the tokenizer, processor (for multimodal models), and model based on configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PreTrainedModel, PreTrainedTokenizer, PeftConfig | None, ProcessorMixin | None]\nTuple containing model, tokenizer, peft_config (if LoRA / QLoRA, else None), and processor (if multimodal, else None).\n\n\n\n\n\n\n\ntrain.setup_model_and_trainer(cfg, dataset_meta)\nLoad model, tokenizer, trainer, etc. Helper function to encapsulate the full\ntrainer setup.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration dictionary with training parameters.\nrequired\n\n\ndataset_meta\nTrainDatasetMeta\nObject with training, validation datasets and metadata.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[HFRLTrainerBuilder | HFCausalTrainerBuilder, PeftModel | PreTrainedModel, PreTrainedTokenizer, PeftConfig | None, ProcessorMixin | None]\nTuple of: - Trainer (Causal or RLHF) - Model - Tokenizer - PEFT config - Processor\n\n\n\n\n\n\n\ntrain.setup_model_card(cfg)\nSet up the Axolotl badge and add the Axolotl config to the model card if available.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\ntrain.setup_reference_model(cfg, tokenizer)\nSet up the reference model for RL training if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntokenizer\nPreTrainedTokenizer\nThe tokenizer to use for the reference model.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nPreTrainedModel | None\nReference model if needed for RL training, None otherwise.\n\n\n\n\n\n\n\ntrain.setup_signal_handler(cfg, model, safe_serialization)\nSet up signal handler for graceful termination.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\nmodel\nPreTrainedModel\nThe model to save on termination\nrequired\n\n\nsafe_serialization\nbool\nWhether to use safe serialization when saving\nrequired\n\n\n\n\n\n\n\ntrain.train(cfg, dataset_meta)\nTrain a model on the given dataset.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration dictionary with training parameters\nrequired\n\n\ndataset_meta\nTrainDatasetMeta\nObject with training, validation datasets and metadata\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PeftModel | PreTrainedModel, PreTrainedTokenizer, Trainer]\nTuple of (model, tokenizer) after training"
  },
  {
    "objectID": "docs/api/train.html#functions",
    "href": "docs/api/train.html#functions",
    "title": "train",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ncreate_model_card\nCreate a model card for the trained model if needed.\n\n\ndetermine_resume_checkpoint\nDetermine the checkpoint to resume from based on configuration.\n\n\nexecute_training\nExecute the training process with appropriate SDP kernel configurations.\n\n\nhandle_untrained_tokens_fix\nApply fixes for untrained tokens if configured.\n\n\nsave_initial_configs\nSave initial configurations before training.\n\n\nsave_trained_model\nSave the trained model according to configuration and training setup.\n\n\nsetup_model_and_tokenizer\nLoad the tokenizer, processor (for multimodal models), and model based on configuration.\n\n\nsetup_model_and_trainer\nLoad model, tokenizer, trainer, etc. Helper function to encapsulate the full\n\n\nsetup_model_card\nSet up the Axolotl badge and add the Axolotl config to the model card if available.\n\n\nsetup_reference_model\nSet up the reference model for RL training if needed.\n\n\nsetup_signal_handler\nSet up signal handler for graceful termination.\n\n\ntrain\nTrain a model on the given dataset.\n\n\n\n\n\ntrain.create_model_card(cfg, trainer)\nCreate a model card for the trained model if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object with model card creation capabilities.\nrequired\n\n\n\n\n\n\n\ntrain.determine_resume_checkpoint(cfg)\nDetermine the checkpoint to resume from based on configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr | None\nPath to the checkpoint to resume from, or None if not resuming.\n\n\n\n\n\n\n\ntrain.execute_training(cfg, trainer, resume_from_checkpoint)\nExecute the training process with appropriate SDP kernel configurations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntrainer\nAny\nThe configured trainer object.\nrequired\n\n\nresume_from_checkpoint\nstr | None\nPath to checkpoint to resume from, if applicable.\nrequired\n\n\n\n\n\n\n\ntrain.handle_untrained_tokens_fix(\n    cfg,\n    model,\n    tokenizer,\n    train_dataset,\n    safe_serialization,\n)\nApply fixes for untrained tokens if configured.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\nmodel\nPreTrainedModel\nThe model to apply fixes to.\nrequired\n\n\ntokenizer\nPreTrainedTokenizer\nThe tokenizer for token identification.\nrequired\n\n\ntrain_dataset\nDataset\nThe training dataset to use.\nrequired\n\n\nsafe_serialization\nbool\nWhether to use safe serialization when saving.\nrequired\n\n\n\n\n\n\n\ntrain.save_initial_configs(cfg, tokenizer, model, peft_config, processor)\nSave initial configurations before training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntokenizer\nPreTrainedTokenizer\nThe tokenizer to save.\nrequired\n\n\nmodel\nPreTrainedModel\nThe model to save configuration for.\nrequired\n\n\npeft_config\nPeftConfig | None\nThe PEFT configuration to save if applicable.\nrequired\n\n\n\n\n\n\n\ntrain.save_trained_model(cfg, trainer, model, safe_serialization)\nSave the trained model according to configuration and training setup.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntrainer\nAny\nThe trainer object.\nrequired\n\n\nmodel\nPreTrainedModel\nThe trained model to save.\nrequired\n\n\nsafe_serialization\nbool\nWhether to use safe serialization.\nrequired\n\n\n\n\n\n\n\ntrain.setup_model_and_tokenizer(cfg)\nLoad the tokenizer, processor (for multimodal models), and model based on configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PreTrainedModel, PreTrainedTokenizer, PeftConfig | None, ProcessorMixin | None]\nTuple containing model, tokenizer, peft_config (if LoRA / QLoRA, else None), and processor (if multimodal, else None).\n\n\n\n\n\n\n\ntrain.setup_model_and_trainer(cfg, dataset_meta)\nLoad model, tokenizer, trainer, etc. Helper function to encapsulate the full\ntrainer setup.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration dictionary with training parameters.\nrequired\n\n\ndataset_meta\nTrainDatasetMeta\nObject with training, validation datasets and metadata.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[HFRLTrainerBuilder | HFCausalTrainerBuilder, PeftModel | PreTrainedModel, PreTrainedTokenizer, PeftConfig | None, ProcessorMixin | None]\nTuple of: - Trainer (Causal or RLHF) - Model - Tokenizer - PEFT config - Processor\n\n\n\n\n\n\n\ntrain.setup_model_card(cfg)\nSet up the Axolotl badge and add the Axolotl config to the model card if available.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\ntrain.setup_reference_model(cfg, tokenizer)\nSet up the reference model for RL training if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ntokenizer\nPreTrainedTokenizer\nThe tokenizer to use for the reference model.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nPreTrainedModel | None\nReference model if needed for RL training, None otherwise.\n\n\n\n\n\n\n\ntrain.setup_signal_handler(cfg, model, safe_serialization)\nSet up signal handler for graceful termination.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\nmodel\nPreTrainedModel\nThe model to save on termination\nrequired\n\n\nsafe_serialization\nbool\nWhether to use safe serialization when saving\nrequired\n\n\n\n\n\n\n\ntrain.train(cfg, dataset_meta)\nTrain a model on the given dataset.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration dictionary with training parameters\nrequired\n\n\ndataset_meta\nTrainDatasetMeta\nObject with training, validation datasets and metadata\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PeftModel | PreTrainedModel, PreTrainedTokenizer, Trainer]\nTuple of (model, tokenizer) after training"
  },
  {
    "objectID": "docs/api/cli.preprocess.html",
    "href": "docs/api/cli.preprocess.html",
    "title": "cli.preprocess",
    "section": "",
    "text": "cli.preprocess\nCLI to run preprocessing of a dataset.\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_preprocess.\n\n\ndo_preprocess\nPreprocesses dataset specified in axolotl config.\n\n\n\n\n\ncli.preprocess.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_preprocess.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.preprocess.do_preprocess(cfg, cli_args)\nPreprocesses dataset specified in axolotl config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nPreprocessCliArgs\nPreprocessing-specific CLI arguments.\nrequired"
  },
  {
    "objectID": "docs/api/cli.preprocess.html#functions",
    "href": "docs/api/cli.preprocess.html#functions",
    "title": "cli.preprocess",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_preprocess.\n\n\ndo_preprocess\nPreprocesses dataset specified in axolotl config.\n\n\n\n\n\ncli.preprocess.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_preprocess.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.preprocess.do_preprocess(cfg, cli_args)\nPreprocesses dataset specified in axolotl config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nPreprocessCliArgs\nPreprocessing-specific CLI arguments.\nrequired"
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.zephyr.html",
    "href": "docs/api/prompt_strategies.dpo.zephyr.html",
    "title": "prompt_strategies.dpo.zephyr",
    "section": "",
    "text": "prompt_strategies.dpo.zephyr\nprompt_strategies.dpo.zephyr\nDPO strategies for zephyr"
  },
  {
    "objectID": "docs/api/core.training_args.html",
    "href": "docs/api/core.training_args.html",
    "title": "core.training_args",
    "section": "",
    "text": "core.training_args\nextra axolotl specific training args\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlCPOConfig\nCPO config for CPO training\n\n\nAxolotlKTOConfig\nKTO config for KTO training\n\n\nAxolotlORPOConfig\nORPO config for ORPO training\n\n\nAxolotlPRMConfig\nPRM config for PRM training\n\n\nAxolotlRewardConfig\nReward config for Reward training\n\n\nAxolotlTrainingArguments\nTraining arguments for Causal trainer\n\n\nAxolotlTrainingMixins\nMixin class for the Axolotl training args.\n\n\n\n\n\ncore.training_args.AxolotlCPOConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n    simpo_gamma=None,\n)\nCPO config for CPO training\n\n\n\ncore.training_args.AxolotlKTOConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nKTO config for KTO training\n\n\n\ncore.training_args.AxolotlORPOConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nORPO config for ORPO training\n\n\n\ncore.training_args.AxolotlPRMConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nPRM config for PRM training\n\n\n\ncore.training_args.AxolotlRewardConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nReward config for Reward training\n\n\n\ncore.training_args.AxolotlTrainingArguments(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nTraining arguments for Causal trainer\nThis code is duplicated due to HF TrainingArguments not setting output_dir with a\ndefault value so it can’t be used as a mixin.\n\n\n\ncore.training_args.AxolotlTrainingMixins(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nMixin class for the Axolotl training args."
  },
  {
    "objectID": "docs/api/core.training_args.html#classes",
    "href": "docs/api/core.training_args.html#classes",
    "title": "core.training_args",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlCPOConfig\nCPO config for CPO training\n\n\nAxolotlKTOConfig\nKTO config for KTO training\n\n\nAxolotlORPOConfig\nORPO config for ORPO training\n\n\nAxolotlPRMConfig\nPRM config for PRM training\n\n\nAxolotlRewardConfig\nReward config for Reward training\n\n\nAxolotlTrainingArguments\nTraining arguments for Causal trainer\n\n\nAxolotlTrainingMixins\nMixin class for the Axolotl training args.\n\n\n\n\n\ncore.training_args.AxolotlCPOConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n    simpo_gamma=None,\n)\nCPO config for CPO training\n\n\n\ncore.training_args.AxolotlKTOConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nKTO config for KTO training\n\n\n\ncore.training_args.AxolotlORPOConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nORPO config for ORPO training\n\n\n\ncore.training_args.AxolotlPRMConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nPRM config for PRM training\n\n\n\ncore.training_args.AxolotlRewardConfig(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nReward config for Reward training\n\n\n\ncore.training_args.AxolotlTrainingArguments(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nTraining arguments for Causal trainer\nThis code is duplicated due to HF TrainingArguments not setting output_dir with a\ndefault value so it can’t be used as a mixin.\n\n\n\ncore.training_args.AxolotlTrainingMixins(\n    model_type=None,\n    lr_quadratic_warmup=False,\n    pretraining=False,\n    sample_packing=False,\n    sample_packing_sequentially=False,\n    multipack_real_batches=False,\n    eval_sample_packing=None,\n    sample_packing_efficiency=1.0,\n    sample_packing_bin_size=200,\n    sample_packing_group_size=100000,\n    max_seq_length=2048,\n    relora_steps=None,\n    relora_warmup_steps=None,\n    relora_anneal_steps=None,\n    relora_prune_ratio=0.9,\n    bench_split='eval',\n    bench_dataset='pharaouk/dharma-1/dharma_1_mini.json',\n    do_bench_eval=False,\n    do_causal_lm_eval=False,\n    max_bench_samples=None,\n    bench_source_max_len=2048,\n    dataloader_prefetch_factor=None,\n    cosine_min_lr_ratio=None,\n    cosine_constant_lr_ratio=None,\n    loraplus_lr_ratio=None,\n    loraplus_lr_embedding=1e-06,\n    embedding_lr_scale=None,\n    lr_groups=None,\n    embedding_lr=None,\n    qlora=False,\n    orpo_alpha=None,\n    lisa_n_layers=None,\n    lisa_step_interval=None,\n    lisa_layers_attribute=None,\n    curriculum_sampling=None,\n    alternate_lr_scheduler_type=None,\n    chat_template=None,\n    kd_ce_alpha=None,\n    kd_alpha=1.0,\n    kd_temperature=1.0,\n    kd_zscore_base_temp=None,\n    kd_top_k_before_softmax=None,\n    adam_beta3=None,\n    adam_epsilon2=None,\n    image_size=None,\n    image_resize_algorithm=None,\n)\nMixin class for the Axolotl training args."
  },
  {
    "objectID": "docs/api/datasets.html",
    "href": "docs/api/datasets.html",
    "title": "datasets",
    "section": "",
    "text": "datasets\nModule containing Dataset functionality\n\n\n\n\n\nName\nDescription\n\n\n\n\nConstantLengthDataset\nIterable dataset that returns constant length chunks of tokens from stream of text files.\n\n\nTokenizedPromptDataset\nDataset that returns tokenized prompts from a stream of text files.\n\n\n\n\n\ndatasets.ConstantLengthDataset(tokenizer, datasets, seq_length=2048)\nIterable dataset that returns constant length chunks of tokens from stream of text files.\nArgs:\ntokenizer (Tokenizer): The processor used for processing the data.\ndataset (dataset.Dataset): Dataset with text files.\nseq_length (int): Length of token sequences to return.\n\n\n\ndatasets.TokenizedPromptDataset(\n    prompt_tokenizer,\n    dataset,\n    process_count=None,\n    keep_in_memory=False,\n    **kwargs,\n)\nDataset that returns tokenized prompts from a stream of text files.\nArgs:\nprompt_tokenizer (PromptTokenizingStrategy): The prompt tokenizing method for processing the data.\ndataset (dataset.Dataset): Dataset with text files.\nprocess_count (int): Number of processes to use for tokenizing.\nkeep_in_memory (bool): Whether to keep the tokenized dataset in memory."
  },
  {
    "objectID": "docs/api/datasets.html#classes",
    "href": "docs/api/datasets.html#classes",
    "title": "datasets",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nConstantLengthDataset\nIterable dataset that returns constant length chunks of tokens from stream of text files.\n\n\nTokenizedPromptDataset\nDataset that returns tokenized prompts from a stream of text files.\n\n\n\n\n\ndatasets.ConstantLengthDataset(tokenizer, datasets, seq_length=2048)\nIterable dataset that returns constant length chunks of tokens from stream of text files.\nArgs:\ntokenizer (Tokenizer): The processor used for processing the data.\ndataset (dataset.Dataset): Dataset with text files.\nseq_length (int): Length of token sequences to return.\n\n\n\ndatasets.TokenizedPromptDataset(\n    prompt_tokenizer,\n    dataset,\n    process_count=None,\n    keep_in_memory=False,\n    **kwargs,\n)\nDataset that returns tokenized prompts from a stream of text files.\nArgs:\nprompt_tokenizer (PromptTokenizingStrategy): The prompt tokenizing method for processing the data.\ndataset (dataset.Dataset): Dataset with text files.\nprocess_count (int): Number of processes to use for tokenizing.\nkeep_in_memory (bool): Whether to keep the tokenized dataset in memory."
  },
  {
    "objectID": "docs/api/models.mamba.modeling_mamba.html",
    "href": "docs/api/models.mamba.modeling_mamba.html",
    "title": "models.mamba.modeling_mamba",
    "section": "",
    "text": "models.mamba.modeling_mamba\nmodels.mamba.modeling_mamba"
  },
  {
    "objectID": "docs/api/kernels.utils.html",
    "href": "docs/api/kernels.utils.html",
    "title": "kernels.utils",
    "section": "",
    "text": "kernels.utils\nkernels.utils\nUtilities for axolotl.kernels submodules."
  },
  {
    "objectID": "docs/api/loaders.adapter.html",
    "href": "docs/api/loaders.adapter.html",
    "title": "loaders.adapter",
    "section": "",
    "text": "loaders.adapter\nAdapter loading functionality, including LoRA / QLoRA and associated utils\n\n\n\n\n\nName\nDescription\n\n\n\n\nsetup_quantized_meta_for_peft\nReplaces quant_state.to with a dummy function to prevent PEFT from moving quant_state to meta device\n\n\nsetup_quantized_peft_meta_for_training\nReplaces dummy quant_state.to method with the original function to allow training to continue\n\n\n\n\n\nloaders.adapter.setup_quantized_meta_for_peft(model)\nReplaces quant_state.to with a dummy function to prevent PEFT from moving quant_state to meta device\n\n\n\nloaders.adapter.setup_quantized_peft_meta_for_training(model)\nReplaces dummy quant_state.to method with the original function to allow training to continue"
  },
  {
    "objectID": "docs/api/loaders.adapter.html#functions",
    "href": "docs/api/loaders.adapter.html#functions",
    "title": "loaders.adapter",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nsetup_quantized_meta_for_peft\nReplaces quant_state.to with a dummy function to prevent PEFT from moving quant_state to meta device\n\n\nsetup_quantized_peft_meta_for_training\nReplaces dummy quant_state.to method with the original function to allow training to continue\n\n\n\n\n\nloaders.adapter.setup_quantized_meta_for_peft(model)\nReplaces quant_state.to with a dummy function to prevent PEFT from moving quant_state to meta device\n\n\n\nloaders.adapter.setup_quantized_peft_meta_for_training(model)\nReplaces dummy quant_state.to method with the original function to allow training to continue"
  },
  {
    "objectID": "docs/api/core.chat.format.chatml.html",
    "href": "docs/api/core.chat.format.chatml.html",
    "title": "core.chat.format.chatml",
    "section": "",
    "text": "core.chat.format.chatml\ncore.chat.format.chatml\nChatML transformation functions for MessageContents"
  },
  {
    "objectID": "docs/api/core.trainers.dpo.trainer.html",
    "href": "docs/api/core.trainers.dpo.trainer.html",
    "title": "core.trainers.dpo.trainer",
    "section": "",
    "text": "core.trainers.dpo.trainer\nDPO trainer for axolotl\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlDPOTrainer\nExtend the base DPOTrainer for axolotl helpers.\n\n\n\n\n\ncore.trainers.dpo.trainer.AxolotlDPOTrainer(*args, dataset_tags=None, **kwargs)\nExtend the base DPOTrainer for axolotl helpers.\n\n\n\n\n\nName\nDescription\n\n\n\n\npush_to_hub\nOverwrite the push_to_hub method in order to force-add the tags when pushing\n\n\n\n\n\ncore.trainers.dpo.trainer.AxolotlDPOTrainer.push_to_hub(*args, **kwargs)\nOverwrite the push_to_hub method in order to force-add the tags when pushing\nthe model on the Hub. Please refer to ~transformers.Trainer.push_to_hub\nfor more details."
  },
  {
    "objectID": "docs/api/core.trainers.dpo.trainer.html#classes",
    "href": "docs/api/core.trainers.dpo.trainer.html#classes",
    "title": "core.trainers.dpo.trainer",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlDPOTrainer\nExtend the base DPOTrainer for axolotl helpers.\n\n\n\n\n\ncore.trainers.dpo.trainer.AxolotlDPOTrainer(*args, dataset_tags=None, **kwargs)\nExtend the base DPOTrainer for axolotl helpers.\n\n\n\n\n\nName\nDescription\n\n\n\n\npush_to_hub\nOverwrite the push_to_hub method in order to force-add the tags when pushing\n\n\n\n\n\ncore.trainers.dpo.trainer.AxolotlDPOTrainer.push_to_hub(*args, **kwargs)\nOverwrite the push_to_hub method in order to force-add the tags when pushing\nthe model on the Hub. Please refer to ~transformers.Trainer.push_to_hub\nfor more details."
  },
  {
    "objectID": "docs/api/cli.cloud.modal_.html",
    "href": "docs/api/cli.cloud.modal_.html",
    "title": "cli.cloud.modal_",
    "section": "",
    "text": "cli.cloud.modal_\nModal Cloud support from CLI\n\n\n\n\n\nName\nDescription\n\n\n\n\nModalCloud\nModal Cloud implementation.\n\n\n\n\n\ncli.cloud.modal_.ModalCloud(config, app=None)\nModal Cloud implementation.\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nrun_cmd\nRun a command inside a folder, with Modal Volume reloading before and commit on success.\n\n\n\n\n\ncli.cloud.modal_.run_cmd(cmd, run_folder, volumes=None)\nRun a command inside a folder, with Modal Volume reloading before and commit on success."
  },
  {
    "objectID": "docs/api/cli.cloud.modal_.html#classes",
    "href": "docs/api/cli.cloud.modal_.html#classes",
    "title": "cli.cloud.modal_",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nModalCloud\nModal Cloud implementation.\n\n\n\n\n\ncli.cloud.modal_.ModalCloud(config, app=None)\nModal Cloud implementation."
  },
  {
    "objectID": "docs/api/cli.cloud.modal_.html#functions",
    "href": "docs/api/cli.cloud.modal_.html#functions",
    "title": "cli.cloud.modal_",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nrun_cmd\nRun a command inside a folder, with Modal Volume reloading before and commit on success.\n\n\n\n\n\ncli.cloud.modal_.run_cmd(cmd, run_folder, volumes=None)\nRun a command inside a folder, with Modal Volume reloading before and commit on success."
  },
  {
    "objectID": "docs/api/prompt_strategies.stepwise_supervised.html",
    "href": "docs/api/prompt_strategies.stepwise_supervised.html",
    "title": "prompt_strategies.stepwise_supervised",
    "section": "",
    "text": "prompt_strategies.stepwise_supervised\nModule for stepwise datasets, typically including a prompt and reasoning traces,\nand (optionally) per-step, or per-prompt-trace labels for reward modelling.\n\n\n\n\n\nName\nDescription\n\n\n\n\nStepwiseSupervisedPromptTokenizingStrategy\nTokenizing strategy for supervised stepwise datasets, typically used for COT-reasoning.\n\n\n\n\n\nprompt_strategies.stepwise_supervised.StepwiseSupervisedPromptTokenizingStrategy(\n    tokenizer,\n    sequence_len=2048,\n    step_separator='\\n',\n    max_completion_length=None,\n    train_on_last_step_only=False,\n)\nTokenizing strategy for supervised stepwise datasets, typically used for COT-reasoning.\nThese datasets should include the following columns:\n- prompt: the prompt text\n- completions: a list of n completion steps\n- labels: a list of n labels indicating the “correctness” of each step"
  },
  {
    "objectID": "docs/api/prompt_strategies.stepwise_supervised.html#classes",
    "href": "docs/api/prompt_strategies.stepwise_supervised.html#classes",
    "title": "prompt_strategies.stepwise_supervised",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nStepwiseSupervisedPromptTokenizingStrategy\nTokenizing strategy for supervised stepwise datasets, typically used for COT-reasoning.\n\n\n\n\n\nprompt_strategies.stepwise_supervised.StepwiseSupervisedPromptTokenizingStrategy(\n    tokenizer,\n    sequence_len=2048,\n    step_separator='\\n',\n    max_completion_length=None,\n    train_on_last_step_only=False,\n)\nTokenizing strategy for supervised stepwise datasets, typically used for COT-reasoning.\nThese datasets should include the following columns:\n- prompt: the prompt text\n- completions: a list of n completion steps\n- labels: a list of n labels indicating the “correctness” of each step"
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.passthrough.html",
    "href": "docs/api/prompt_strategies.dpo.passthrough.html",
    "title": "prompt_strategies.dpo.passthrough",
    "section": "",
    "text": "prompt_strategies.dpo.passthrough\nprompt_strategies.dpo.passthrough\nDPO prompt strategies passthrough/zero-processing strategy"
  },
  {
    "objectID": "docs/api/monkeypatch.relora.html",
    "href": "docs/api/monkeypatch.relora.html",
    "title": "monkeypatch.relora",
    "section": "",
    "text": "monkeypatch.relora\nImplements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune.\n\n\n\n\n\nName\nDescription\n\n\n\n\nReLoRACallback\nCallback to merge LoRA weights into the base model and save full-weight checkpoints\n\n\nReLoRAScheduler\nWraps another scheduler to apply per-lora-restart learning rate warmups.\n\n\n\n\n\nmonkeypatch.relora.ReLoRACallback(cfg)\nCallback to merge LoRA weights into the base model and save full-weight checkpoints\n\n\n\nmonkeypatch.relora.ReLoRAScheduler(\n    optimizer,\n    inner_schedule,\n    relora_steps,\n    warmup_steps,\n    anneal_steps=1,\n    min_lr_scale=0.001,\n)\nWraps another scheduler to apply per-lora-restart learning rate warmups."
  },
  {
    "objectID": "docs/api/monkeypatch.relora.html#classes",
    "href": "docs/api/monkeypatch.relora.html#classes",
    "title": "monkeypatch.relora",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nReLoRACallback\nCallback to merge LoRA weights into the base model and save full-weight checkpoints\n\n\nReLoRAScheduler\nWraps another scheduler to apply per-lora-restart learning rate warmups.\n\n\n\n\n\nmonkeypatch.relora.ReLoRACallback(cfg)\nCallback to merge LoRA weights into the base model and save full-weight checkpoints\n\n\n\nmonkeypatch.relora.ReLoRAScheduler(\n    optimizer,\n    inner_schedule,\n    relora_steps,\n    warmup_steps,\n    anneal_steps=1,\n    min_lr_scale=0.001,\n)\nWraps another scheduler to apply per-lora-restart learning rate warmups."
  },
  {
    "objectID": "docs/api/kernels.swiglu.html",
    "href": "docs/api/kernels.swiglu.html",
    "title": "kernels.swiglu",
    "section": "",
    "text": "kernels.swiglu\nModule for definition of SwiGLU Triton kernels.\nSee “GLU Variants Improve Transformer” (https://arxiv.org/abs/2002.05202).\nCredit to unsloth (https://unsloth.ai/) for inspiration for this implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nswiglu_backward\nSwiGLU backward pass using in-place operations.\n\n\nswiglu_forward\nSwiGLU forward pass. Computes SwiGLU activation: x * sigmoid(x) * up, where\n\n\n\n\n\nkernels.swiglu.swiglu_backward(grad_output, gate, up)\nSwiGLU backward pass using in-place operations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to output, shape [batch, seq_len, hidden_dim].\nrequired\n\n\ngate\ntorch.Tensor\nGate tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple containing: - Forward pass output (h) - Gradient with respect to gate (df) - Gradient with respect to up-projection (de)\n\n\n\n\n\n\n\nkernels.swiglu.swiglu_forward(gate, up)\nSwiGLU forward pass. Computes SwiGLU activation: x * sigmoid(x) * up, where\nx is the gate tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngate\ntorch.Tensor\nInput gate tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor of shape [batch, seq_len, hidden_dim]."
  },
  {
    "objectID": "docs/api/kernels.swiglu.html#functions",
    "href": "docs/api/kernels.swiglu.html#functions",
    "title": "kernels.swiglu",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nswiglu_backward\nSwiGLU backward pass using in-place operations.\n\n\nswiglu_forward\nSwiGLU forward pass. Computes SwiGLU activation: x * sigmoid(x) * up, where\n\n\n\n\n\nkernels.swiglu.swiglu_backward(grad_output, gate, up)\nSwiGLU backward pass using in-place operations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to output, shape [batch, seq_len, hidden_dim].\nrequired\n\n\ngate\ntorch.Tensor\nGate tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor from forward pass, shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple containing: - Forward pass output (h) - Gradient with respect to gate (df) - Gradient with respect to up-projection (de)\n\n\n\n\n\n\n\nkernels.swiglu.swiglu_forward(gate, up)\nSwiGLU forward pass. Computes SwiGLU activation: x * sigmoid(x) * up, where\nx is the gate tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ngate\ntorch.Tensor\nInput gate tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\nup\ntorch.Tensor\nUp-projection tensor of shape [batch, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor of shape [batch, seq_len, hidden_dim]."
  },
  {
    "objectID": "docs/api/cli.cloud.base.html",
    "href": "docs/api/cli.cloud.base.html",
    "title": "cli.cloud.base",
    "section": "",
    "text": "cli.cloud.base\nbase class for cloud platforms from cli\n\n\n\n\n\nName\nDescription\n\n\n\n\nCloud\nAbstract base class for cloud platforms.\n\n\n\n\n\ncli.cloud.base.Cloud()\nAbstract base class for cloud platforms."
  },
  {
    "objectID": "docs/api/cli.cloud.base.html#classes",
    "href": "docs/api/cli.cloud.base.html#classes",
    "title": "cli.cloud.base",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nCloud\nAbstract base class for cloud platforms.\n\n\n\n\n\ncli.cloud.base.Cloud()\nAbstract base class for cloud platforms."
  },
  {
    "objectID": "docs/api/common.datasets.html",
    "href": "docs/api/common.datasets.html",
    "title": "common.datasets",
    "section": "",
    "text": "common.datasets\nDataset loading utilities.\n\n\n\n\n\nName\nDescription\n\n\n\n\nTrainDatasetMeta\nDataclass with fields for training and validation datasets and metadata.\n\n\n\n\n\ncommon.datasets.TrainDatasetMeta(\n    train_dataset,\n    eval_dataset=None,\n    total_num_steps=None,\n)\nDataclass with fields for training and validation datasets and metadata.\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nload_datasets\nLoads one or more training or evaluation datasets, calling\n\n\nload_preference_datasets\nLoads one or more training or evaluation datasets for RL training using paired\n\n\nsample_dataset\nRandomly sample num_samples samples from dataset.\n\n\n\n\n\ncommon.datasets.load_datasets(cfg, cli_args=None, debug=False)\nLoads one or more training or evaluation datasets, calling\naxolotl.utils.data.prepare_dataset. Optionally, logs out debug information.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nPreprocessCliArgs | TrainerCliArgs | None\nCommand-specific CLI arguments.\nNone\n\n\ndebug\nbool\nWhether to print out tokenization of sample\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainDatasetMeta\nDataclass with fields for training and evaluation datasets and the computed\n\n\n\nTrainDatasetMeta\ntotal_num_steps.\n\n\n\n\n\n\n\ncommon.datasets.load_preference_datasets(cfg, cli_args)\nLoads one or more training or evaluation datasets for RL training using paired\npreference data, calling axolotl.utils.data.rl.load_prepare_preference_datasets.\nOptionally, logs out debug information.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nUnion[PreprocessCliArgs, TrainerCliArgs]\nCommand-specific CLI arguments.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainDatasetMeta\nDataclass with fields for training and evaluation datasets and the computed\n\n\n\nTrainDatasetMeta\ntotal_num_steps.\n\n\n\n\n\n\n\ncommon.datasets.sample_dataset(dataset, num_samples)\nRandomly sample num_samples samples from dataset.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndataset\nDataset\nDataset.\nrequired\n\n\nnum_samples\nint\nNumber of samples to return.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nDataset\nRandom sample (with replacement) of examples in dataset."
  },
  {
    "objectID": "docs/api/common.datasets.html#classes",
    "href": "docs/api/common.datasets.html#classes",
    "title": "common.datasets",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nTrainDatasetMeta\nDataclass with fields for training and validation datasets and metadata.\n\n\n\n\n\ncommon.datasets.TrainDatasetMeta(\n    train_dataset,\n    eval_dataset=None,\n    total_num_steps=None,\n)\nDataclass with fields for training and validation datasets and metadata."
  },
  {
    "objectID": "docs/api/common.datasets.html#functions",
    "href": "docs/api/common.datasets.html#functions",
    "title": "common.datasets",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nload_datasets\nLoads one or more training or evaluation datasets, calling\n\n\nload_preference_datasets\nLoads one or more training or evaluation datasets for RL training using paired\n\n\nsample_dataset\nRandomly sample num_samples samples from dataset.\n\n\n\n\n\ncommon.datasets.load_datasets(cfg, cli_args=None, debug=False)\nLoads one or more training or evaluation datasets, calling\naxolotl.utils.data.prepare_dataset. Optionally, logs out debug information.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nPreprocessCliArgs | TrainerCliArgs | None\nCommand-specific CLI arguments.\nNone\n\n\ndebug\nbool\nWhether to print out tokenization of sample\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainDatasetMeta\nDataclass with fields for training and evaluation datasets and the computed\n\n\n\nTrainDatasetMeta\ntotal_num_steps.\n\n\n\n\n\n\n\ncommon.datasets.load_preference_datasets(cfg, cli_args)\nLoads one or more training or evaluation datasets for RL training using paired\npreference data, calling axolotl.utils.data.rl.load_prepare_preference_datasets.\nOptionally, logs out debug information.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nUnion[PreprocessCliArgs, TrainerCliArgs]\nCommand-specific CLI arguments.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainDatasetMeta\nDataclass with fields for training and evaluation datasets and the computed\n\n\n\nTrainDatasetMeta\ntotal_num_steps.\n\n\n\n\n\n\n\ncommon.datasets.sample_dataset(dataset, num_samples)\nRandomly sample num_samples samples from dataset.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndataset\nDataset\nDataset.\nrequired\n\n\nnum_samples\nint\nNumber of samples to return.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nDataset\nRandom sample (with replacement) of examples in dataset."
  },
  {
    "objectID": "docs/api/core.trainers.grpo.sampler.html",
    "href": "docs/api/core.trainers.grpo.sampler.html",
    "title": "core.trainers.grpo.sampler",
    "section": "",
    "text": "core.trainers.grpo.sampler\nRepeat random sampler (similar to the one implemented in\nhttps://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py) that adds\nsequence parallelism functionality; i.e., duplicating data across ranks in the same\nsequence parallel group.\n\n\n\n\n\nName\nDescription\n\n\n\n\nSequenceParallelRepeatRandomSampler\nSampler for GRPO training with sequence parallelism.\n\n\n\n\n\ncore.trainers.grpo.sampler.SequenceParallelRepeatRandomSampler(\n    dataset,\n    mini_repeat_count,\n    world_size,\n    rank,\n    batch_size=1,\n    repeat_count=1,\n    sequence_parallel_degree=1,\n    shuffle=True,\n    seed=0,\n    drop_last=False,\n)\nSampler for GRPO training with sequence parallelism.\nThis sampler ensures:\n- Ranks in the same sequence parallel (SP) group receive identical data.\n- Each index is repeated multiple times for sampling different completions.\n- Entire batches are repeated for reuse in multiple updates.\n- Data is properly distributed across SP groups.\nIn the table below, the values represent dataset indices. Each SP group has\nsequence_parallel_degree = 2 GPUs working together on the same data. There are 2\nSP groups (SP0 and SP1), with world_size = 4 total GPUs.\n                                       Sequence Parallel Groups\n                                |       SP0        |       SP1        |\n                                |  GPU 0  |  GPU 1 |  GPU 2  |  GPU 3 |\n            global_step  step    <---> mini_repeat_count=3\n                                    <----------> batch_size=2 per SP group\ngrad_accum=2 ▲ ▲ 0 0 [0 0 0 1 1 1] [2 2 2 3 3 3] <- SP groups get different data\n▼ | 0 1 [0 0 0 1 1 1] [2 2 2 3 3 3] <- Same data for each SP group GPU\n|\n| 1 2 [0 0 0 1 1 1] [2 2 2 3 3 3] <- Repeat same indices for iterations\nnum_iterations=2 ▼ 1 3 [0 0 0 1 1 1] [2 2 2 3 3 3] <- When using gradient accumulation\n                 2       4         [4 4 4  5 5 5]     [6 6 6  7 7 7]   <- New batch of data indices\n                 2       5         [4 4 4  5 5 5]     [6 6 6  7 7 7]\n                                    ...\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndataset\nSized\nDataset to sample from.\nrequired\n\n\nmini_repeat_count\nint\nHow many times to repeat each sample immediately.\nrequired\n\n\nworld_size\nint\nTotal number of processes.\nrequired\n\n\nrank\nint\nRank of current process.\nrequired\n\n\nbatch_size\nint\nNumber of samples per batch.\n1\n\n\nrepeat_count\nint\nHow many times to repeat the full sampling process.\n1\n\n\nsequence_parallel_degree\nint\nNumber of ranks in a sequence parallel group.\n1\n\n\nshuffle\nbool\nWhether to shuffle the dataset.\nTrue\n\n\nseed\nint\nRandom seed for shuffling.\n0\n\n\ndrop_last\nbool\nWhether to drop the last incomplete batch.\nFalse\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nset_epoch\nSets the epoch for this sampler.\n\n\n\n\n\ncore.trainers.grpo.sampler.SequenceParallelRepeatRandomSampler.set_epoch(epoch)\nSets the epoch for this sampler.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nepoch\nint\nEpoch number to use for shuffling.\nrequired"
  },
  {
    "objectID": "docs/api/core.trainers.grpo.sampler.html#classes",
    "href": "docs/api/core.trainers.grpo.sampler.html#classes",
    "title": "core.trainers.grpo.sampler",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nSequenceParallelRepeatRandomSampler\nSampler for GRPO training with sequence parallelism.\n\n\n\n\n\ncore.trainers.grpo.sampler.SequenceParallelRepeatRandomSampler(\n    dataset,\n    mini_repeat_count,\n    world_size,\n    rank,\n    batch_size=1,\n    repeat_count=1,\n    sequence_parallel_degree=1,\n    shuffle=True,\n    seed=0,\n    drop_last=False,\n)\nSampler for GRPO training with sequence parallelism.\nThis sampler ensures:\n- Ranks in the same sequence parallel (SP) group receive identical data.\n- Each index is repeated multiple times for sampling different completions.\n- Entire batches are repeated for reuse in multiple updates.\n- Data is properly distributed across SP groups.\nIn the table below, the values represent dataset indices. Each SP group has\nsequence_parallel_degree = 2 GPUs working together on the same data. There are 2\nSP groups (SP0 and SP1), with world_size = 4 total GPUs.\n                                       Sequence Parallel Groups\n                                |       SP0        |       SP1        |\n                                |  GPU 0  |  GPU 1 |  GPU 2  |  GPU 3 |\n            global_step  step    <---> mini_repeat_count=3\n                                    <----------> batch_size=2 per SP group\ngrad_accum=2 ▲ ▲ 0 0 [0 0 0 1 1 1] [2 2 2 3 3 3] <- SP groups get different data\n▼ | 0 1 [0 0 0 1 1 1] [2 2 2 3 3 3] <- Same data for each SP group GPU\n|\n| 1 2 [0 0 0 1 1 1] [2 2 2 3 3 3] <- Repeat same indices for iterations\nnum_iterations=2 ▼ 1 3 [0 0 0 1 1 1] [2 2 2 3 3 3] <- When using gradient accumulation\n                 2       4         [4 4 4  5 5 5]     [6 6 6  7 7 7]   <- New batch of data indices\n                 2       5         [4 4 4  5 5 5]     [6 6 6  7 7 7]\n                                    ...\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndataset\nSized\nDataset to sample from.\nrequired\n\n\nmini_repeat_count\nint\nHow many times to repeat each sample immediately.\nrequired\n\n\nworld_size\nint\nTotal number of processes.\nrequired\n\n\nrank\nint\nRank of current process.\nrequired\n\n\nbatch_size\nint\nNumber of samples per batch.\n1\n\n\nrepeat_count\nint\nHow many times to repeat the full sampling process.\n1\n\n\nsequence_parallel_degree\nint\nNumber of ranks in a sequence parallel group.\n1\n\n\nshuffle\nbool\nWhether to shuffle the dataset.\nTrue\n\n\nseed\nint\nRandom seed for shuffling.\n0\n\n\ndrop_last\nbool\nWhether to drop the last incomplete batch.\nFalse\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nset_epoch\nSets the epoch for this sampler.\n\n\n\n\n\ncore.trainers.grpo.sampler.SequenceParallelRepeatRandomSampler.set_epoch(epoch)\nSets the epoch for this sampler.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nepoch\nint\nEpoch number to use for shuffling.\nrequired"
  },
  {
    "objectID": "docs/api/cli.merge_lora.html",
    "href": "docs/api/cli.merge_lora.html",
    "title": "cli.merge_lora",
    "section": "",
    "text": "cli.merge_lora\nCLI to merge a trained LoRA into a base model.\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_merge_lora. Note that various\n\n\ndo_merge_lora\nCalls transformers’ merge_and_unload on the model given in the axolotl config\n\n\n\n\n\ncli.merge_lora.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_merge_lora. Note that various\nconfig values will be overwritten to allow the LoRA merge logic to work as expected\n(load_in_8bit=False, load_in4bit=False, flash_attention=False, etc.).\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf target directory for LoRA merged model does not exist.\n\n\n\n\n\n\n\ncli.merge_lora.do_merge_lora(cfg)\nCalls transformers’ merge_and_unload on the model given in the axolotl config\nalong with the LoRA adapters to combine them into a single base model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired"
  },
  {
    "objectID": "docs/api/cli.merge_lora.html#functions",
    "href": "docs/api/cli.merge_lora.html#functions",
    "title": "cli.merge_lora",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_merge_lora. Note that various\n\n\ndo_merge_lora\nCalls transformers’ merge_and_unload on the model given in the axolotl config\n\n\n\n\n\ncli.merge_lora.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_merge_lora. Note that various\nconfig values will be overwritten to allow the LoRA merge logic to work as expected\n(load_in_8bit=False, load_in4bit=False, flash_attention=False, etc.).\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf target directory for LoRA merged model does not exist.\n\n\n\n\n\n\n\ncli.merge_lora.do_merge_lora(cfg)\nCalls transformers’ merge_and_unload on the model given in the axolotl config\nalong with the LoRA adapters to combine them into a single base model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired"
  },
  {
    "objectID": "docs/api/utils.model_shard_quant.html",
    "href": "docs/api/utils.model_shard_quant.html",
    "title": "utils.model_shard_quant",
    "section": "",
    "text": "utils.model_shard_quant\nmodule to handle loading model on cpu/meta device for FSDP\n\n\n\n\n\nName\nDescription\n\n\n\n\nload_and_quantize\nLoads value tensor into submodule of module, optionally skipping skip_names and converting to dtype.\n\n\n\n\n\nutils.model_shard_quant.load_and_quantize(\n    module,\n    name,\n    value,\n    device=None,\n    dtype=None,\n    skip_names=None,\n    to_cpu=False,\n    to_meta=False,\n    verbose=False,\n    quant_method='bnb',\n)\nLoads value tensor into submodule of module, optionally skipping skip_names and converting to dtype.\nQuantizes Params4bit on device then places on “cpu” if to_cpu=True or “meta” if to_meta=True."
  },
  {
    "objectID": "docs/api/utils.model_shard_quant.html#functions",
    "href": "docs/api/utils.model_shard_quant.html#functions",
    "title": "utils.model_shard_quant",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nload_and_quantize\nLoads value tensor into submodule of module, optionally skipping skip_names and converting to dtype.\n\n\n\n\n\nutils.model_shard_quant.load_and_quantize(\n    module,\n    name,\n    value,\n    device=None,\n    dtype=None,\n    skip_names=None,\n    to_cpu=False,\n    to_meta=False,\n    verbose=False,\n    quant_method='bnb',\n)\nLoads value tensor into submodule of module, optionally skipping skip_names and converting to dtype.\nQuantizes Params4bit on device then places on “cpu” if to_cpu=True or “meta” if to_meta=True."
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.llama3.html",
    "href": "docs/api/prompt_strategies.dpo.llama3.html",
    "title": "prompt_strategies.dpo.llama3",
    "section": "",
    "text": "prompt_strategies.dpo.llama3\nDPO strategies for llama-3 chat template\n\n\n\n\n\nName\nDescription\n\n\n\n\nargilla_chat\nfor argilla/dpo-mix-7k conversations\n\n\nicr\nchatml transforms for datasets with system, input, chosen, rejected\n\n\nintel\nFor Intel Orca DPO Pairs\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.dpo.llama3.argilla_chat(cfg, **kwargs)\nfor argilla/dpo-mix-7k conversations\n\n\n\nprompt_strategies.dpo.llama3.icr(cfg, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected\nex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs\n\n\n\nprompt_strategies.dpo.llama3.intel(cfg, **kwargs)\nFor Intel Orca DPO Pairs\n\n\n\nprompt_strategies.dpo.llama3.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations"
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.llama3.html#functions",
    "href": "docs/api/prompt_strategies.dpo.llama3.html#functions",
    "title": "prompt_strategies.dpo.llama3",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nargilla_chat\nfor argilla/dpo-mix-7k conversations\n\n\nicr\nchatml transforms for datasets with system, input, chosen, rejected\n\n\nintel\nFor Intel Orca DPO Pairs\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.dpo.llama3.argilla_chat(cfg, **kwargs)\nfor argilla/dpo-mix-7k conversations\n\n\n\nprompt_strategies.dpo.llama3.icr(cfg, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected\nex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs\n\n\n\nprompt_strategies.dpo.llama3.intel(cfg, **kwargs)\nFor Intel Orca DPO Pairs\n\n\n\nprompt_strategies.dpo.llama3.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations"
  },
  {
    "objectID": "docs/api/integrations.liger.args.html",
    "href": "docs/api/integrations.liger.args.html",
    "title": "integrations.liger.args",
    "section": "",
    "text": "integrations.liger.args\nModule for handling LIGER input arguments.\n\n\n\n\n\nName\nDescription\n\n\n\n\nLigerArgs\nInput args for LIGER.\n\n\n\n\n\nintegrations.liger.args.LigerArgs()\nInput args for LIGER."
  },
  {
    "objectID": "docs/api/integrations.liger.args.html#classes",
    "href": "docs/api/integrations.liger.args.html#classes",
    "title": "integrations.liger.args",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nLigerArgs\nInput args for LIGER.\n\n\n\n\n\nintegrations.liger.args.LigerArgs()\nInput args for LIGER."
  },
  {
    "objectID": "docs/api/utils.callbacks.profiler.html",
    "href": "docs/api/utils.callbacks.profiler.html",
    "title": "utils.callbacks.profiler",
    "section": "",
    "text": "utils.callbacks.profiler\nHF Trainer callback for creating pytorch profiling snapshots\n\n\n\n\n\nName\nDescription\n\n\n\n\nPytorchProfilerCallback\nPyTorch Profiler callback to create snapshots of GPU memory usage at specified steps.\n\n\n\n\n\nutils.callbacks.profiler.PytorchProfilerCallback(steps_to_profile=5)\nPyTorch Profiler callback to create snapshots of GPU memory usage at specified steps."
  },
  {
    "objectID": "docs/api/utils.callbacks.profiler.html#classes",
    "href": "docs/api/utils.callbacks.profiler.html#classes",
    "title": "utils.callbacks.profiler",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nPytorchProfilerCallback\nPyTorch Profiler callback to create snapshots of GPU memory usage at specified steps.\n\n\n\n\n\nutils.callbacks.profiler.PytorchProfilerCallback(steps_to_profile=5)\nPyTorch Profiler callback to create snapshots of GPU memory usage at specified steps."
  },
  {
    "objectID": "docs/api/utils.chat_templates.html",
    "href": "docs/api/utils.chat_templates.html",
    "title": "utils.chat_templates",
    "section": "",
    "text": "utils.chat_templates\nThis module provides functionality for selecting chat templates based on user choices.\nThese templates are used for formatting messages in a conversation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_chat_template\nFinds the correct chat_template based on the user’s choice, jinja_template, and tokenizer.\n\n\nregister_chat_template\nRegisters chat templates.\n\n\n\n\n\nutils.chat_templates.get_chat_template(\n    user_choice,\n    jinja_template=None,\n    tokenizer=None,\n)\nFinds the correct chat_template based on the user’s choice, jinja_template, and tokenizer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nuser_choice\nstr\nThe user’s choice of template.\nrequired\n\n\njinja_template\nOptional[str]\nThe jinja template string. Defaults to None.\nNone\n\n\ntokenizer\nOptional[PreTrainedTokenizerBase]\nThe tokenizer. Defaults to None.\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nstr\nstr\nThe chosen template string.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the user_choice is not found in the templates.\n\n\n\n\n\n\n\nutils.chat_templates.register_chat_template(template_name, chat_template)\nRegisters chat templates.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntemplate_name\nstr\nThe name of the template.\nrequired\n\n\nchat_template\nstr\nThe template string.\nrequired"
  },
  {
    "objectID": "docs/api/utils.chat_templates.html#functions",
    "href": "docs/api/utils.chat_templates.html#functions",
    "title": "utils.chat_templates",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nget_chat_template\nFinds the correct chat_template based on the user’s choice, jinja_template, and tokenizer.\n\n\nregister_chat_template\nRegisters chat templates.\n\n\n\n\n\nutils.chat_templates.get_chat_template(\n    user_choice,\n    jinja_template=None,\n    tokenizer=None,\n)\nFinds the correct chat_template based on the user’s choice, jinja_template, and tokenizer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nuser_choice\nstr\nThe user’s choice of template.\nrequired\n\n\njinja_template\nOptional[str]\nThe jinja template string. Defaults to None.\nNone\n\n\ntokenizer\nOptional[PreTrainedTokenizerBase]\nThe tokenizer. Defaults to None.\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nstr\nstr\nThe chosen template string.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the user_choice is not found in the templates.\n\n\n\n\n\n\n\nutils.chat_templates.register_chat_template(template_name, chat_template)\nRegisters chat templates.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntemplate_name\nstr\nThe name of the template.\nrequired\n\n\nchat_template\nstr\nThe template string.\nrequired"
  },
  {
    "objectID": "docs/api/core.builders.base.html",
    "href": "docs/api/core.builders.base.html",
    "title": "core.builders.base",
    "section": "",
    "text": "core.builders.base\nBase class for trainer builder\n\n\n\n\n\nName\nDescription\n\n\n\n\nTrainerBuilderBase\nBase class for trainer builder.\n\n\n\n\n\ncore.builders.base.TrainerBuilderBase(cfg, model, tokenizer, processor=None)\nBase class for trainer builder.\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_post_trainer_create_callbacks\nCallbacks added after the trainer is created, usually b/c these need access to the trainer\n\n\n\n\n\ncore.builders.base.TrainerBuilderBase.get_post_trainer_create_callbacks(trainer)\nCallbacks added after the trainer is created, usually b/c these need access to the trainer"
  },
  {
    "objectID": "docs/api/core.builders.base.html#classes",
    "href": "docs/api/core.builders.base.html#classes",
    "title": "core.builders.base",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nTrainerBuilderBase\nBase class for trainer builder.\n\n\n\n\n\ncore.builders.base.TrainerBuilderBase(cfg, model, tokenizer, processor=None)\nBase class for trainer builder.\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_post_trainer_create_callbacks\nCallbacks added after the trainer is created, usually b/c these need access to the trainer\n\n\n\n\n\ncore.builders.base.TrainerBuilderBase.get_post_trainer_create_callbacks(trainer)\nCallbacks added after the trainer is created, usually b/c these need access to the trainer"
  },
  {
    "objectID": "docs/api/core.trainers.mamba.html",
    "href": "docs/api/core.trainers.mamba.html",
    "title": "core.trainers.mamba",
    "section": "",
    "text": "core.trainers.mamba\nModule for mamba trainer\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlMambaTrainer\nMamba specific trainer to handle loss calculation\n\n\n\n\n\ncore.trainers.mamba.AxolotlMambaTrainer(\n    *_args,\n    bench_data_collator=None,\n    eval_data_collator=None,\n    dataset_tags=None,\n    **kwargs,\n)\nMamba specific trainer to handle loss calculation"
  },
  {
    "objectID": "docs/api/core.trainers.mamba.html#classes",
    "href": "docs/api/core.trainers.mamba.html#classes",
    "title": "core.trainers.mamba",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlMambaTrainer\nMamba specific trainer to handle loss calculation\n\n\n\n\n\ncore.trainers.mamba.AxolotlMambaTrainer(\n    *_args,\n    bench_data_collator=None,\n    eval_data_collator=None,\n    dataset_tags=None,\n    **kwargs,\n)\nMamba specific trainer to handle loss calculation"
  },
  {
    "objectID": "docs/api/prompt_strategies.orpo.chat_template.html",
    "href": "docs/api/prompt_strategies.orpo.chat_template.html",
    "title": "prompt_strategies.orpo.chat_template",
    "section": "",
    "text": "prompt_strategies.orpo.chat_template\nchatml prompt tokenization strategy for ORPO\n\n\n\n\n\nName\nDescription\n\n\n\n\nMessage\nmessage/turn\n\n\nMessageList\nconversation\n\n\nORPODatasetParsingStrategy\nStrategy to parse chosen rejected dataset into messagelist\n\n\nORPOPrompter\nSingle Turn prompter for ORPO\n\n\nORPOTokenizingStrategy\nrejected_input_ids\n\n\n\n\n\nprompt_strategies.orpo.chat_template.Message()\nmessage/turn\n\n\n\nprompt_strategies.orpo.chat_template.MessageList()\nconversation\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy()\nStrategy to parse chosen rejected dataset into messagelist\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_chosen_conversation_thread\nDataset structure mappings\n\n\nget_prompt\nMap the data to extract everything up to the last turn\n\n\nget_rejected_conversation_thread\nDataset structure mappings\n\n\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy.get_chosen_conversation_thread(\n    prompt,\n)\nDataset structure mappings\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy.get_prompt(\n    prompt,\n)\nMap the data to extract everything up to the last turn\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy.get_rejected_conversation_thread(\n    prompt,\n)\nDataset structure mappings\n\n\n\n\n\nprompt_strategies.orpo.chat_template.ORPOPrompter(chat_template, tokenizer)\nSingle Turn prompter for ORPO\n\n\n\nprompt_strategies.orpo.chat_template.ORPOTokenizingStrategy(\n    *args,\n    dataset_parser=None,\n    **kwargs,\n)\nrejected_input_ids\ninput_ids\nrejected_attention_mask\nattention_mask\nrejected_labels\nlabels\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nload\nchatml transforms for datasets with system, input, chosen, rejected\n\n\n\n\n\nprompt_strategies.orpo.chat_template.load(tokenizer, cfg, ds_cfg=None, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected"
  },
  {
    "objectID": "docs/api/prompt_strategies.orpo.chat_template.html#classes",
    "href": "docs/api/prompt_strategies.orpo.chat_template.html#classes",
    "title": "prompt_strategies.orpo.chat_template",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMessage\nmessage/turn\n\n\nMessageList\nconversation\n\n\nORPODatasetParsingStrategy\nStrategy to parse chosen rejected dataset into messagelist\n\n\nORPOPrompter\nSingle Turn prompter for ORPO\n\n\nORPOTokenizingStrategy\nrejected_input_ids\n\n\n\n\n\nprompt_strategies.orpo.chat_template.Message()\nmessage/turn\n\n\n\nprompt_strategies.orpo.chat_template.MessageList()\nconversation\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy()\nStrategy to parse chosen rejected dataset into messagelist\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_chosen_conversation_thread\nDataset structure mappings\n\n\nget_prompt\nMap the data to extract everything up to the last turn\n\n\nget_rejected_conversation_thread\nDataset structure mappings\n\n\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy.get_chosen_conversation_thread(\n    prompt,\n)\nDataset structure mappings\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy.get_prompt(\n    prompt,\n)\nMap the data to extract everything up to the last turn\n\n\n\nprompt_strategies.orpo.chat_template.ORPODatasetParsingStrategy.get_rejected_conversation_thread(\n    prompt,\n)\nDataset structure mappings\n\n\n\n\n\nprompt_strategies.orpo.chat_template.ORPOPrompter(chat_template, tokenizer)\nSingle Turn prompter for ORPO\n\n\n\nprompt_strategies.orpo.chat_template.ORPOTokenizingStrategy(\n    *args,\n    dataset_parser=None,\n    **kwargs,\n)\nrejected_input_ids\ninput_ids\nrejected_attention_mask\nattention_mask\nrejected_labels\nlabels"
  },
  {
    "objectID": "docs/api/prompt_strategies.orpo.chat_template.html#functions",
    "href": "docs/api/prompt_strategies.orpo.chat_template.html#functions",
    "title": "prompt_strategies.orpo.chat_template",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nload\nchatml transforms for datasets with system, input, chosen, rejected\n\n\n\n\n\nprompt_strategies.orpo.chat_template.load(tokenizer, cfg, ds_cfg=None, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected"
  },
  {
    "objectID": "docs/api/utils.data.pretraining.html",
    "href": "docs/api/utils.data.pretraining.html",
    "title": "utils.data.pretraining",
    "section": "",
    "text": "utils.data.pretraining\nutils.data.pretraining\ndata handling specific to pretraining"
  },
  {
    "objectID": "docs/api/utils.schemas.peft.html",
    "href": "docs/api/utils.schemas.peft.html",
    "title": "utils.schemas.peft",
    "section": "",
    "text": "utils.schemas.peft\nPydantic models for PEFT-related configuration\n\n\n\n\n\nName\nDescription\n\n\n\n\nLoftQConfig\nLoftQ configuration subset\n\n\nLoraConfig\nPeft / LoRA configuration subset\n\n\nPeftConfig\npeftq configuration subset\n\n\nReLoRAConfig\nReLoRA configuration subset\n\n\n\n\n\nutils.schemas.peft.LoftQConfig()\nLoftQ configuration subset\n\n\n\nutils.schemas.peft.LoraConfig()\nPeft / LoRA configuration subset\n\n\n\nutils.schemas.peft.PeftConfig()\npeftq configuration subset\n\n\n\nutils.schemas.peft.ReLoRAConfig()\nReLoRA configuration subset"
  },
  {
    "objectID": "docs/api/utils.schemas.peft.html#classes",
    "href": "docs/api/utils.schemas.peft.html#classes",
    "title": "utils.schemas.peft",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nLoftQConfig\nLoftQ configuration subset\n\n\nLoraConfig\nPeft / LoRA configuration subset\n\n\nPeftConfig\npeftq configuration subset\n\n\nReLoRAConfig\nReLoRA configuration subset\n\n\n\n\n\nutils.schemas.peft.LoftQConfig()\nLoftQ configuration subset\n\n\n\nutils.schemas.peft.LoraConfig()\nPeft / LoRA configuration subset\n\n\n\nutils.schemas.peft.PeftConfig()\npeftq configuration subset\n\n\n\nutils.schemas.peft.ReLoRAConfig()\nReLoRA configuration subset"
  },
  {
    "objectID": "docs/api/monkeypatch.trainer_fsdp_optim.html",
    "href": "docs/api/monkeypatch.trainer_fsdp_optim.html",
    "title": "monkeypatch.trainer_fsdp_optim",
    "section": "",
    "text": "monkeypatch.trainer_fsdp_optim\nfix for FSDP optimizer save in trainer w 4.47.0\n\n\n\n\n\nName\nDescription\n\n\n\n\npatch_training_loop_for_fsdp\nmonkeypatch for fixing the training loop for fsdp with optimizer save\n\n\n\n\n\nmonkeypatch.trainer_fsdp_optim.patch_training_loop_for_fsdp()\nmonkeypatch for fixing the training loop for fsdp with optimizer save"
  },
  {
    "objectID": "docs/api/monkeypatch.trainer_fsdp_optim.html#functions",
    "href": "docs/api/monkeypatch.trainer_fsdp_optim.html#functions",
    "title": "monkeypatch.trainer_fsdp_optim",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\npatch_training_loop_for_fsdp\nmonkeypatch for fixing the training loop for fsdp with optimizer save\n\n\n\n\n\nmonkeypatch.trainer_fsdp_optim.patch_training_loop_for_fsdp()\nmonkeypatch for fixing the training loop for fsdp with optimizer save"
  },
  {
    "objectID": "docs/api/prompt_strategies.chat_template.html",
    "href": "docs/api/prompt_strategies.chat_template.html",
    "title": "prompt_strategies.chat_template",
    "section": "",
    "text": "prompt_strategies.chat_template\nHF Chat Templates prompt strategy\n\n\n\n\n\nName\nDescription\n\n\n\n\nChatTemplatePrompter\nPrompter for HF chat templates\n\n\nChatTemplateStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nStrategyLoader\nLoad chat template strategy based on configuration.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter(\n    tokenizer,\n    chat_template,\n    processor=None,\n    max_length=2048,\n    message_property_mappings=None,\n    message_field_training=None,\n    message_field_training_detail=None,\n    field_messages='messages',\n    field_system='system',\n    roles=None,\n    drop_system_message=False,\n)\nPrompter for HF chat templates\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs,\n    sequence_len,\n    roles_to_train=None,\n    train_on_eos=None,\n    train_on_eot=None,\n    eot_tokens=None,\n    split_thinking=False,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\nfind_turn\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\ntokenize_prompt\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_first_eot_token(\n    input_ids,\n    start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_turn(turns, turn_idx)\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.tokenize_prompt(prompt)\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.StrategyLoader()\nLoad chat template strategy based on configuration."
  },
  {
    "objectID": "docs/api/prompt_strategies.chat_template.html#classes",
    "href": "docs/api/prompt_strategies.chat_template.html#classes",
    "title": "prompt_strategies.chat_template",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nChatTemplatePrompter\nPrompter for HF chat templates\n\n\nChatTemplateStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nStrategyLoader\nLoad chat template strategy based on configuration.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter(\n    tokenizer,\n    chat_template,\n    processor=None,\n    max_length=2048,\n    message_property_mappings=None,\n    message_field_training=None,\n    message_field_training_detail=None,\n    field_messages='messages',\n    field_system='system',\n    roles=None,\n    drop_system_message=False,\n)\nPrompter for HF chat templates\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs,\n    sequence_len,\n    roles_to_train=None,\n    train_on_eos=None,\n    train_on_eot=None,\n    eot_tokens=None,\n    split_thinking=False,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\nfind_turn\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\ntokenize_prompt\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_first_eot_token(\n    input_ids,\n    start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_turn(turns, turn_idx)\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.tokenize_prompt(prompt)\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.StrategyLoader()\nLoad chat template strategy based on configuration."
  },
  {
    "objectID": "docs/api/prompt_strategies.kto.chatml.html",
    "href": "docs/api/prompt_strategies.kto.chatml.html",
    "title": "prompt_strategies.kto.chatml",
    "section": "",
    "text": "prompt_strategies.kto.chatml\nKTO strategies for chatml\n\n\n\n\n\nName\nDescription\n\n\n\n\nargilla_chat\nfor argilla/kto-mix-15k conversations\n\n\nintel\nFor Intel Orca KTO\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.kto.chatml.argilla_chat(cfg, **kwargs)\nfor argilla/kto-mix-15k conversations\n\n\n\nprompt_strategies.kto.chatml.intel(cfg, **kwargs)\nFor Intel Orca KTO\nex: argilla/distilabel-intel-orca-kto\n\n\n\nprompt_strategies.kto.chatml.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations\nex: argilla/ultrafeedback-binarized-preferences-cleaned-kto"
  },
  {
    "objectID": "docs/api/prompt_strategies.kto.chatml.html#functions",
    "href": "docs/api/prompt_strategies.kto.chatml.html#functions",
    "title": "prompt_strategies.kto.chatml",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nargilla_chat\nfor argilla/kto-mix-15k conversations\n\n\nintel\nFor Intel Orca KTO\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.kto.chatml.argilla_chat(cfg, **kwargs)\nfor argilla/kto-mix-15k conversations\n\n\n\nprompt_strategies.kto.chatml.intel(cfg, **kwargs)\nFor Intel Orca KTO\nex: argilla/distilabel-intel-orca-kto\n\n\n\nprompt_strategies.kto.chatml.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations\nex: argilla/ultrafeedback-binarized-preferences-cleaned-kto"
  },
  {
    "objectID": "docs/api/monkeypatch.btlm_attn_hijack_flash.html",
    "href": "docs/api/monkeypatch.btlm_attn_hijack_flash.html",
    "title": "monkeypatch.btlm_attn_hijack_flash",
    "section": "",
    "text": "monkeypatch.btlm_attn_hijack_flash\nmonkeypatch.btlm_attn_hijack_flash\nFlash attention monkey patch for cerebras btlm model"
  },
  {
    "objectID": "docs/api/core.chat.format.shared.html",
    "href": "docs/api/core.chat.format.shared.html",
    "title": "core.chat.format.shared",
    "section": "",
    "text": "core.chat.format.shared\ncore.chat.format.shared\nshared functions for format transforms"
  },
  {
    "objectID": "docs/api/logging_config.html",
    "href": "docs/api/logging_config.html",
    "title": "logging_config",
    "section": "",
    "text": "logging_config\nCommon logging module for axolotl\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlLogger\nA Logger that automatically rejects non-axolotl INFOs.\n\n\nAxolotlOrWarnErrorFilter\nAllows ANY WARNING or higher (unless overridden by LOG_LEVEL)\n\n\nColorfulFormatter\nFormatter to add coloring to log messages by log type\n\n\n\n\n\nlogging_config.AxolotlLogger(name, level=logging.NOTSET)\nA Logger that automatically rejects non-axolotl INFOs.\n\n\n\nlogging_config.AxolotlOrWarnErrorFilter(**kwargs)\nAllows ANY WARNING or higher (unless overridden by LOG_LEVEL)\nAllows axolotl.* at INFO or higher (unless overridden by AXOLOTL_LOG_LEVEL)\nDrops all other records (i.e. non-axolotl.INFO, DEBUG, etc. by default)\n\n\n\nlogging_config.ColorfulFormatter()\nFormatter to add coloring to log messages by log type\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nconfigure_logging\nConfigure with default logging\n\n\n\n\n\nlogging_config.configure_logging()\nConfigure with default logging"
  },
  {
    "objectID": "docs/api/logging_config.html#classes",
    "href": "docs/api/logging_config.html#classes",
    "title": "logging_config",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlLogger\nA Logger that automatically rejects non-axolotl INFOs.\n\n\nAxolotlOrWarnErrorFilter\nAllows ANY WARNING or higher (unless overridden by LOG_LEVEL)\n\n\nColorfulFormatter\nFormatter to add coloring to log messages by log type\n\n\n\n\n\nlogging_config.AxolotlLogger(name, level=logging.NOTSET)\nA Logger that automatically rejects non-axolotl INFOs.\n\n\n\nlogging_config.AxolotlOrWarnErrorFilter(**kwargs)\nAllows ANY WARNING or higher (unless overridden by LOG_LEVEL)\nAllows axolotl.* at INFO or higher (unless overridden by AXOLOTL_LOG_LEVEL)\nDrops all other records (i.e. non-axolotl.INFO, DEBUG, etc. by default)\n\n\n\nlogging_config.ColorfulFormatter()\nFormatter to add coloring to log messages by log type"
  },
  {
    "objectID": "docs/api/logging_config.html#functions",
    "href": "docs/api/logging_config.html#functions",
    "title": "logging_config",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nconfigure_logging\nConfigure with default logging\n\n\n\n\n\nlogging_config.configure_logging()\nConfigure with default logging"
  },
  {
    "objectID": "docs/api/cli.sweeps.html",
    "href": "docs/api/cli.sweeps.html",
    "title": "cli.sweeps",
    "section": "",
    "text": "cli.sweeps\nUtilities for handling sweeps over configs for axolotl train CLI command\n\n\n\n\n\nName\nDescription\n\n\n\n\ngenerate_sweep_configs\nRecursively generates all possible configurations by applying sweeps to the base config.\n\n\n\n\n\ncli.sweeps.generate_sweep_configs(base_config, sweeps_config)\nRecursively generates all possible configurations by applying sweeps to the base config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbase_config\ndict\nThe original configuration dictionary\nrequired\n\n\nsweeps_config\ndict\nDictionary where keys are parameters and values are either: - lists of values to sweep independently - or for paired values, a list of dicts under the ’_’ key\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nlist\nlist[dict[str, list]]\nList of all possible configuration dictionaries\n\n\n\n\n\n\nsweeps_config = {\n‘learning_rate’: [0.1, 0.01],\n’_’: [\n{‘load_in_8bit’: True, ‘adapter’: ‘lora’},\n{‘load_in_4bit’: True, ‘adapter’: ‘qlora’}\n]\n}"
  },
  {
    "objectID": "docs/api/cli.sweeps.html#functions",
    "href": "docs/api/cli.sweeps.html#functions",
    "title": "cli.sweeps",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ngenerate_sweep_configs\nRecursively generates all possible configurations by applying sweeps to the base config.\n\n\n\n\n\ncli.sweeps.generate_sweep_configs(base_config, sweeps_config)\nRecursively generates all possible configurations by applying sweeps to the base config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbase_config\ndict\nThe original configuration dictionary\nrequired\n\n\nsweeps_config\ndict\nDictionary where keys are parameters and values are either: - lists of values to sweep independently - or for paired values, a list of dicts under the ’_’ key\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nlist\nlist[dict[str, list]]\nList of all possible configuration dictionaries\n\n\n\n\n\n\nsweeps_config = {\n‘learning_rate’: [0.1, 0.01],\n’_’: [\n{‘load_in_8bit’: True, ‘adapter’: ‘lora’},\n{‘load_in_4bit’: True, ‘adapter’: ‘qlora’}\n]\n}"
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.chat_template.html",
    "href": "docs/api/prompt_strategies.dpo.chat_template.html",
    "title": "prompt_strategies.dpo.chat_template",
    "section": "",
    "text": "prompt_strategies.dpo.chat_template\nprompt_strategies.dpo.chat_template\nDPO prompt strategies for using tokenizer chat templates."
  },
  {
    "objectID": "docs/api/prompt_tokenizers.html",
    "href": "docs/api/prompt_tokenizers.html",
    "title": "prompt_tokenizers",
    "section": "",
    "text": "prompt_tokenizers\nModule containing PromptTokenizingStrategy and Prompter classes\n\n\n\n\n\nName\nDescription\n\n\n\n\nAlpacaMultipleChoicePromptTokenizingStrategy\nTokenizing strategy for Alpaca Multiple Choice prompts.\n\n\nAlpacaPromptTokenizingStrategy\nTokenizing strategy for Alpaca prompts.\n\n\nAlpacaReflectionPTStrategy\nTokenizing strategy for Alpaca Reflection prompts.\n\n\nDatasetWrappingStrategy\nAbstract class for wrapping datasets for Chat Messages\n\n\nGPTeacherPromptTokenizingStrategy\nTokenizing strategy for GPTeacher prompts.\n\n\nInstructionPromptTokenizingStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nInvalidDataException\nException raised when the data is invalid\n\n\nJeopardyPromptTokenizingStrategy\nTokenizing strategy for Jeopardy prompts.\n\n\nNomicGPT4AllPromptTokenizingStrategy\nTokenizing strategy for NomicGPT4All prompts.\n\n\nOpenAssistantPromptTokenizingStrategy\nTokenizing strategy for OpenAssistant prompts.\n\n\nPromptTokenizingStrategy\nAbstract class for tokenizing strategies\n\n\nReflectionPromptTokenizingStrategy\nTokenizing strategy for Reflection prompts.\n\n\nSummarizeTLDRPromptTokenizingStrategy\nTokenizing strategy for SummarizeTLDR prompts.\n\n\n\n\n\nprompt_tokenizers.AlpacaMultipleChoicePromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Alpaca Multiple Choice prompts.\n\n\n\nprompt_tokenizers.AlpacaPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Alpaca prompts.\n\n\n\nprompt_tokenizers.AlpacaReflectionPTStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Alpaca Reflection prompts.\n\n\n\nprompt_tokenizers.DatasetWrappingStrategy()\nAbstract class for wrapping datasets for Chat Messages\n\n\n\nprompt_tokenizers.GPTeacherPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for GPTeacher prompts.\n\n\n\nprompt_tokenizers.InstructionPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\nprompt_tokenizers.InvalidDataException()\nException raised when the data is invalid\n\n\n\nprompt_tokenizers.JeopardyPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Jeopardy prompts.\n\n\n\nprompt_tokenizers.NomicGPT4AllPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for NomicGPT4All prompts.\n\n\n\nprompt_tokenizers.OpenAssistantPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for OpenAssistant prompts.\n\n\n\nprompt_tokenizers.PromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nAbstract class for tokenizing strategies\n\n\n\nprompt_tokenizers.ReflectionPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Reflection prompts.\n\n\n\nprompt_tokenizers.SummarizeTLDRPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for SummarizeTLDR prompts.\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nparse_tokenized_to_result\nParses the tokenized prompt and append the tokenized input_ids, attention_mask and labels to the result\n\n\ntokenize_prompt_default\nReturns the default values for the tokenize prompt function\n\n\n\n\n\nprompt_tokenizers.parse_tokenized_to_result(\n    result,\n    current_len,\n    res,\n    labels,\n    pad_token_id=None,\n)\nParses the tokenized prompt and append the tokenized input_ids, attention_mask and labels to the result\n\n\n\nprompt_tokenizers.tokenize_prompt_default()\nReturns the default values for the tokenize prompt function"
  },
  {
    "objectID": "docs/api/prompt_tokenizers.html#classes",
    "href": "docs/api/prompt_tokenizers.html#classes",
    "title": "prompt_tokenizers",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAlpacaMultipleChoicePromptTokenizingStrategy\nTokenizing strategy for Alpaca Multiple Choice prompts.\n\n\nAlpacaPromptTokenizingStrategy\nTokenizing strategy for Alpaca prompts.\n\n\nAlpacaReflectionPTStrategy\nTokenizing strategy for Alpaca Reflection prompts.\n\n\nDatasetWrappingStrategy\nAbstract class for wrapping datasets for Chat Messages\n\n\nGPTeacherPromptTokenizingStrategy\nTokenizing strategy for GPTeacher prompts.\n\n\nInstructionPromptTokenizingStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nInvalidDataException\nException raised when the data is invalid\n\n\nJeopardyPromptTokenizingStrategy\nTokenizing strategy for Jeopardy prompts.\n\n\nNomicGPT4AllPromptTokenizingStrategy\nTokenizing strategy for NomicGPT4All prompts.\n\n\nOpenAssistantPromptTokenizingStrategy\nTokenizing strategy for OpenAssistant prompts.\n\n\nPromptTokenizingStrategy\nAbstract class for tokenizing strategies\n\n\nReflectionPromptTokenizingStrategy\nTokenizing strategy for Reflection prompts.\n\n\nSummarizeTLDRPromptTokenizingStrategy\nTokenizing strategy for SummarizeTLDR prompts.\n\n\n\n\n\nprompt_tokenizers.AlpacaMultipleChoicePromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Alpaca Multiple Choice prompts.\n\n\n\nprompt_tokenizers.AlpacaPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Alpaca prompts.\n\n\n\nprompt_tokenizers.AlpacaReflectionPTStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Alpaca Reflection prompts.\n\n\n\nprompt_tokenizers.DatasetWrappingStrategy()\nAbstract class for wrapping datasets for Chat Messages\n\n\n\nprompt_tokenizers.GPTeacherPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for GPTeacher prompts.\n\n\n\nprompt_tokenizers.InstructionPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\nprompt_tokenizers.InvalidDataException()\nException raised when the data is invalid\n\n\n\nprompt_tokenizers.JeopardyPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Jeopardy prompts.\n\n\n\nprompt_tokenizers.NomicGPT4AllPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for NomicGPT4All prompts.\n\n\n\nprompt_tokenizers.OpenAssistantPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for OpenAssistant prompts.\n\n\n\nprompt_tokenizers.PromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nAbstract class for tokenizing strategies\n\n\n\nprompt_tokenizers.ReflectionPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for Reflection prompts.\n\n\n\nprompt_tokenizers.SummarizeTLDRPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for SummarizeTLDR prompts."
  },
  {
    "objectID": "docs/api/prompt_tokenizers.html#functions",
    "href": "docs/api/prompt_tokenizers.html#functions",
    "title": "prompt_tokenizers",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nparse_tokenized_to_result\nParses the tokenized prompt and append the tokenized input_ids, attention_mask and labels to the result\n\n\ntokenize_prompt_default\nReturns the default values for the tokenize prompt function\n\n\n\n\n\nprompt_tokenizers.parse_tokenized_to_result(\n    result,\n    current_len,\n    res,\n    labels,\n    pad_token_id=None,\n)\nParses the tokenized prompt and append the tokenized input_ids, attention_mask and labels to the result\n\n\n\nprompt_tokenizers.tokenize_prompt_default()\nReturns the default values for the tokenize prompt function"
  },
  {
    "objectID": "docs/api/kernels.quantize.html",
    "href": "docs/api/kernels.quantize.html",
    "title": "kernels.quantize",
    "section": "",
    "text": "kernels.quantize\nDequantization utilities for bitsandbytes integration.\n\n\n\n\n\nName\nDescription\n\n\n\n\ndequantize\nFast NF4 dequantization using bitsandbytes CUDA kernels.\n\n\n\n\n\nkernels.quantize.dequantize(W, quant_state=None, out=None)\nFast NF4 dequantization using bitsandbytes CUDA kernels.\nPerforms efficient dequantization of weights from NF4 format using bitsandbytes’\noptimized CUDA implementations. Supports both legacy list and new QuantState\nformats.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nW\ntorch.Tensor\nQuantized weight tensor to dequantize\nrequired\n\n\nquant_state\nQuantState | list | None\nQuantization state containing metadata needed for dequantization. Can be either a QuantState object or legacy list format. If None, returns W unchanged.\nNone\n\n\nout\ntorch.Tensor | None\nOptional output tensor for storing dequantized results. Must match expected shape and dtype if provided.\nNone\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nDequantized tensor in the specified dtype (fp16 or bf16). Will be transposed if\n\n\n\ntorch.Tensor\ninput W was transposed.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf provided output tensor doesn’t match expected shape / dtype.\n\n\n\n\n\n\nUses CUDA streams for better performance when available in newer bitsandbytes\nversions (>0.43.3)."
  },
  {
    "objectID": "docs/api/kernels.quantize.html#functions",
    "href": "docs/api/kernels.quantize.html#functions",
    "title": "kernels.quantize",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndequantize\nFast NF4 dequantization using bitsandbytes CUDA kernels.\n\n\n\n\n\nkernels.quantize.dequantize(W, quant_state=None, out=None)\nFast NF4 dequantization using bitsandbytes CUDA kernels.\nPerforms efficient dequantization of weights from NF4 format using bitsandbytes’\noptimized CUDA implementations. Supports both legacy list and new QuantState\nformats.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nW\ntorch.Tensor\nQuantized weight tensor to dequantize\nrequired\n\n\nquant_state\nQuantState | list | None\nQuantization state containing metadata needed for dequantization. Can be either a QuantState object or legacy list format. If None, returns W unchanged.\nNone\n\n\nout\ntorch.Tensor | None\nOptional output tensor for storing dequantized results. Must match expected shape and dtype if provided.\nNone\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nDequantized tensor in the specified dtype (fp16 or bf16). Will be transposed if\n\n\n\ntorch.Tensor\ninput W was transposed.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf provided output tensor doesn’t match expected shape / dtype.\n\n\n\n\n\n\nUses CUDA streams for better performance when available in newer bitsandbytes\nversions (>0.43.3)."
  },
  {
    "objectID": "docs/api/utils.callbacks.mlflow_.html",
    "href": "docs/api/utils.callbacks.mlflow_.html",
    "title": "utils.callbacks.mlflow_",
    "section": "",
    "text": "utils.callbacks.mlflow_\nMLFlow module for trainer callbacks\n\n\n\n\n\nName\nDescription\n\n\n\n\nSaveAxolotlConfigtoMlflowCallback\nCallback to save axolotl config to mlflow\n\n\n\n\n\nutils.callbacks.mlflow_.SaveAxolotlConfigtoMlflowCallback(axolotl_config_path)\nCallback to save axolotl config to mlflow"
  },
  {
    "objectID": "docs/api/utils.callbacks.mlflow_.html#classes",
    "href": "docs/api/utils.callbacks.mlflow_.html#classes",
    "title": "utils.callbacks.mlflow_",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nSaveAxolotlConfigtoMlflowCallback\nCallback to save axolotl config to mlflow\n\n\n\n\n\nutils.callbacks.mlflow_.SaveAxolotlConfigtoMlflowCallback(axolotl_config_path)\nCallback to save axolotl config to mlflow"
  },
  {
    "objectID": "docs/api/utils.schemas.trl.html",
    "href": "docs/api/utils.schemas.trl.html",
    "title": "utils.schemas.trl",
    "section": "",
    "text": "utils.schemas.trl\nPydantic models for TRL trainer configuration\n\n\n\n\n\nName\nDescription\n\n\n\n\nTRLConfig\nInput args for TRL.\n\n\n\n\n\nutils.schemas.trl.TRLConfig()\nInput args for TRL."
  },
  {
    "objectID": "docs/api/utils.schemas.trl.html#classes",
    "href": "docs/api/utils.schemas.trl.html#classes",
    "title": "utils.schemas.trl",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nTRLConfig\nInput args for TRL.\n\n\n\n\n\nutils.schemas.trl.TRLConfig()\nInput args for TRL."
  },
  {
    "objectID": "docs/api/utils.schemas.config.html",
    "href": "docs/api/utils.schemas.config.html",
    "title": "utils.schemas.config",
    "section": "",
    "text": "utils.schemas.config\nModule with Pydantic models for configuration.\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlConfigWCapabilities\nwrapper to valdiate gpu capabilities with the configured options\n\n\nAxolotlInputConfig\nWrapper of all config options\n\n\n\n\n\nutils.schemas.config.AxolotlConfigWCapabilities()\nwrapper to valdiate gpu capabilities with the configured options\n\n\n\nutils.schemas.config.AxolotlInputConfig()\nWrapper of all config options"
  },
  {
    "objectID": "docs/api/utils.schemas.config.html#classes",
    "href": "docs/api/utils.schemas.config.html#classes",
    "title": "utils.schemas.config",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlConfigWCapabilities\nwrapper to valdiate gpu capabilities with the configured options\n\n\nAxolotlInputConfig\nWrapper of all config options\n\n\n\n\n\nutils.schemas.config.AxolotlConfigWCapabilities()\nwrapper to valdiate gpu capabilities with the configured options\n\n\n\nutils.schemas.config.AxolotlInputConfig()\nWrapper of all config options"
  },
  {
    "objectID": "docs/api/monkeypatch.unsloth_.html",
    "href": "docs/api/monkeypatch.unsloth_.html",
    "title": "monkeypatch.unsloth_",
    "section": "",
    "text": "monkeypatch.unsloth_\nmonkeypatch.unsloth_\nmodule for patching with unsloth optimizations"
  },
  {
    "objectID": "docs/api/cli.checks.html",
    "href": "docs/api/cli.checks.html",
    "title": "cli.checks",
    "section": "",
    "text": "cli.checks\nVarious checks for Axolotl CLI.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncheck_accelerate_default_config\nLogs at warning level if no accelerate config file is found.\n\n\ncheck_user_token\nChecks for HF user info. Check is skipped if HF_HUB_OFFLINE=1.\n\n\n\n\n\ncli.checks.check_accelerate_default_config()\nLogs at warning level if no accelerate config file is found.\n\n\n\ncli.checks.check_user_token()\nChecks for HF user info. Check is skipped if HF_HUB_OFFLINE=1.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nbool\nBoolean indicating successful check (i.e., HF_HUB_OFFLINE=1 or HF user info is retrieved).\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nLocalTokenNotFoundError\nIf HF user info can’t be retrieved."
  },
  {
    "objectID": "docs/api/cli.checks.html#functions",
    "href": "docs/api/cli.checks.html#functions",
    "title": "cli.checks",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ncheck_accelerate_default_config\nLogs at warning level if no accelerate config file is found.\n\n\ncheck_user_token\nChecks for HF user info. Check is skipped if HF_HUB_OFFLINE=1.\n\n\n\n\n\ncli.checks.check_accelerate_default_config()\nLogs at warning level if no accelerate config file is found.\n\n\n\ncli.checks.check_user_token()\nChecks for HF user info. Check is skipped if HF_HUB_OFFLINE=1.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nbool\nBoolean indicating successful check (i.e., HF_HUB_OFFLINE=1 or HF user info is retrieved).\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nLocalTokenNotFoundError\nIf HF user info can’t be retrieved."
  },
  {
    "objectID": "docs/api/core.chat.messages.html",
    "href": "docs/api/core.chat.messages.html",
    "title": "core.chat.messages",
    "section": "",
    "text": "core.chat.messages\ninternal message representations of chat messages\n\n\n\n\n\nName\nDescription\n\n\n\n\nChatFormattedChats\nChat formatted chats with formatter and optional train on inputs\n\n\nChats\ntop level data structure for chat conversations\n\n\nMessageContentTypes\nMessage content types for text, image, audio, tool calls, and tool responses\n\n\nMessageContents\nMessage contents with type, value, metadata, weight, newline, and end of contents\n\n\nMessageRoles\nMessage roles for the system, user, assistant, and tools\n\n\nMessages\nMessages with role, content, metadata, weight, and chat formatting\n\n\nPreferenceChats\nrepresentation for preference data for chat\n\n\nSpecialToken\nSpecial tokens for beginning of string and end of string\n\n\nTool\nTool with description, function, and parameters\n\n\nToolCallContents\nTool call contents with name, arguments, and optional id\n\n\nToolCallFunction\nTool call function with name and arguments\n\n\nToolResponseContents\nTool response contents with name, content, and optional id\n\n\n\n\n\ncore.chat.messages.ChatFormattedChats()\nChat formatted chats with formatter and optional train on inputs\n\n\n\ncore.chat.messages.Chats()\ntop level data structure for chat conversations\n\n\n\ncore.chat.messages.MessageContentTypes()\nMessage content types for text, image, audio, tool calls, and tool responses\n\n\n\ncore.chat.messages.MessageContents()\nMessage contents with type, value, metadata, weight, newline, and end of contents\n\n\n\ncore.chat.messages.MessageRoles()\nMessage roles for the system, user, assistant, and tools\n\n\n\ncore.chat.messages.Messages()\nMessages with role, content, metadata, weight, and chat formatting\n\n\n\ncore.chat.messages.PreferenceChats()\nrepresentation for preference data for chat\n\n\n\ncore.chat.messages.SpecialToken()\nSpecial tokens for beginning of string and end of string\n\n\n\ncore.chat.messages.Tool()\nTool with description, function, and parameters\n\n\n\ncore.chat.messages.ToolCallContents()\nTool call contents with name, arguments, and optional id\n\n\n\ncore.chat.messages.ToolCallFunction()\nTool call function with name and arguments\n\n\n\ncore.chat.messages.ToolResponseContents()\nTool response contents with name, content, and optional id"
  },
  {
    "objectID": "docs/api/core.chat.messages.html#classes",
    "href": "docs/api/core.chat.messages.html#classes",
    "title": "core.chat.messages",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nChatFormattedChats\nChat formatted chats with formatter and optional train on inputs\n\n\nChats\ntop level data structure for chat conversations\n\n\nMessageContentTypes\nMessage content types for text, image, audio, tool calls, and tool responses\n\n\nMessageContents\nMessage contents with type, value, metadata, weight, newline, and end of contents\n\n\nMessageRoles\nMessage roles for the system, user, assistant, and tools\n\n\nMessages\nMessages with role, content, metadata, weight, and chat formatting\n\n\nPreferenceChats\nrepresentation for preference data for chat\n\n\nSpecialToken\nSpecial tokens for beginning of string and end of string\n\n\nTool\nTool with description, function, and parameters\n\n\nToolCallContents\nTool call contents with name, arguments, and optional id\n\n\nToolCallFunction\nTool call function with name and arguments\n\n\nToolResponseContents\nTool response contents with name, content, and optional id\n\n\n\n\n\ncore.chat.messages.ChatFormattedChats()\nChat formatted chats with formatter and optional train on inputs\n\n\n\ncore.chat.messages.Chats()\ntop level data structure for chat conversations\n\n\n\ncore.chat.messages.MessageContentTypes()\nMessage content types for text, image, audio, tool calls, and tool responses\n\n\n\ncore.chat.messages.MessageContents()\nMessage contents with type, value, metadata, weight, newline, and end of contents\n\n\n\ncore.chat.messages.MessageRoles()\nMessage roles for the system, user, assistant, and tools\n\n\n\ncore.chat.messages.Messages()\nMessages with role, content, metadata, weight, and chat formatting\n\n\n\ncore.chat.messages.PreferenceChats()\nrepresentation for preference data for chat\n\n\n\ncore.chat.messages.SpecialToken()\nSpecial tokens for beginning of string and end of string\n\n\n\ncore.chat.messages.Tool()\nTool with description, function, and parameters\n\n\n\ncore.chat.messages.ToolCallContents()\nTool call contents with name, arguments, and optional id\n\n\n\ncore.chat.messages.ToolCallFunction()\nTool call function with name and arguments\n\n\n\ncore.chat.messages.ToolResponseContents()\nTool response contents with name, content, and optional id"
  },
  {
    "objectID": "docs/api/loaders.patch_manager.html",
    "href": "docs/api/loaders.patch_manager.html",
    "title": "loaders.patch_manager",
    "section": "",
    "text": "loaders.patch_manager\nPatch manager class implementation to complement axolotl.loaders.ModelLoader.\nApplies pre- and post-model load patches for various fixes and optimizations.\n\n\n\n\n\nName\nDescription\n\n\n\n\nPatchManager\nManages the application of patches during the model loading process.\n\n\n\n\n\nloaders.patch_manager.PatchManager(cfg, model_config, inference=False)\nManages the application of patches during the model loading process.\n\n\n\n\n\nName\nDescription\n\n\n\n\nhas_flash_attn\nCheck if flash attention is installed.\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_post_model_load_patches\nApply patches that require the model instance.\n\n\napply_pre_model_load_patches\nApply pre-model load patches based on config.\n\n\n\n\n\nloaders.patch_manager.PatchManager.apply_post_model_load_patches(model)\nApply patches that require the model instance.\n\n\n\nloaders.patch_manager.PatchManager.apply_pre_model_load_patches()\nApply pre-model load patches based on config."
  },
  {
    "objectID": "docs/api/loaders.patch_manager.html#classes",
    "href": "docs/api/loaders.patch_manager.html#classes",
    "title": "loaders.patch_manager",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nPatchManager\nManages the application of patches during the model loading process.\n\n\n\n\n\nloaders.patch_manager.PatchManager(cfg, model_config, inference=False)\nManages the application of patches during the model loading process.\n\n\n\n\n\nName\nDescription\n\n\n\n\nhas_flash_attn\nCheck if flash attention is installed.\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_post_model_load_patches\nApply patches that require the model instance.\n\n\napply_pre_model_load_patches\nApply pre-model load patches based on config.\n\n\n\n\n\nloaders.patch_manager.PatchManager.apply_post_model_load_patches(model)\nApply patches that require the model instance.\n\n\n\nloaders.patch_manager.PatchManager.apply_pre_model_load_patches()\nApply pre-model load patches based on config."
  },
  {
    "objectID": "docs/api/monkeypatch.llama_expand_mask.html",
    "href": "docs/api/monkeypatch.llama_expand_mask.html",
    "title": "monkeypatch.llama_expand_mask",
    "section": "",
    "text": "monkeypatch.llama_expand_mask\nmonkeypatch.llama_expand_mask\nexpands the binary attention mask per 3.2.2 of https://arxiv.org/pdf/2107.02027.pdf"
  },
  {
    "objectID": "docs/api/utils.schemas.utils.html",
    "href": "docs/api/utils.schemas.utils.html",
    "title": "utils.schemas.utils",
    "section": "",
    "text": "utils.schemas.utils\nUtilities for Axolotl Pydantic models\n\n\n\n\n\nName\nDescription\n\n\n\n\nhandle_legacy_message_fields_logic\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\n\n\n\n\n\nutils.schemas.utils.handle_legacy_message_fields_logic(data)\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\nPreviously, the config only supported mapping ‘role’ and ‘content’ fields via dedicated config options:\n- message_field_role: Mapped to the role field\n- message_field_content: Mapped to the content field\nThe new system uses message_property_mappings to support arbitrary field mappings:\nmessage_property_mappings:\nrole: source_role_field\ncontent: source_content_field\nadditional_field: source_field\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndata\ndict\nDictionary containing configuration data\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ndict\nUpdated dictionary with message field mappings consolidated\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf there are conflicts between legacy and new mappings"
  },
  {
    "objectID": "docs/api/utils.schemas.utils.html#functions",
    "href": "docs/api/utils.schemas.utils.html#functions",
    "title": "utils.schemas.utils",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nhandle_legacy_message_fields_logic\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\n\n\n\n\n\nutils.schemas.utils.handle_legacy_message_fields_logic(data)\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\nPreviously, the config only supported mapping ‘role’ and ‘content’ fields via dedicated config options:\n- message_field_role: Mapped to the role field\n- message_field_content: Mapped to the content field\nThe new system uses message_property_mappings to support arbitrary field mappings:\nmessage_property_mappings:\nrole: source_role_field\ncontent: source_content_field\nadditional_field: source_field\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndata\ndict\nDictionary containing configuration data\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ndict\nUpdated dictionary with message field mappings consolidated\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf there are conflicts between legacy and new mappings"
  },
  {
    "objectID": "docs/api/integrations.base.html",
    "href": "docs/api/integrations.base.html",
    "title": "integrations.base",
    "section": "",
    "text": "integrations.base\nBase class for all plugins.\nA plugin is a reusable, modular, and self-contained piece of code that extends the functionality of Axolotl.\nPlugins can be used to integrate third-party models, modify the training process, or add new features.\nTo create a new plugin, you need to inherit from the BasePlugin class and implement the required methods.\n\n\n\n\n\nName\nDescription\n\n\n\n\nBaseOptimizerFactory\nBase class for factories to create custom optimizers\n\n\nBasePlugin\nBase class for all plugins. Defines the interface for plugin methods.\n\n\nPluginManager\nThe PluginManager class is responsible for loading and managing plugins. It\n\n\n\n\n\nintegrations.base.BaseOptimizerFactory()\nBase class for factories to create custom optimizers\n\n\n\nintegrations.base.BasePlugin()\nBase class for all plugins. Defines the interface for plugin methods.\nA plugin is a reusable, modular, and self-contained piece of code that extends\nthe functionality of Axolotl. Plugins can be used to integrate third-party models,\nmodify the training process, or add new features.\nTo create a new plugin, you need to inherit from the BasePlugin class and\nimplement the required methods.\n\n\nPlugin methods include:\n- register(cfg): Registers the plugin with the given configuration.\n- load_datasets(cfg): Loads and preprocesses the dataset for training.\n- pre_model_load(cfg): Performs actions before the model is loaded.\n- post_model_build(cfg, model): Performs actions after the model is loaded, but\nbefore LoRA adapters are applied.\n- pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.\n- post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.\n- post_model_load(cfg, model): Performs actions after the model is loaded,\ninclusive of any adapters.\n- post_trainer_create(cfg, trainer): Performs actions after the trainer is\ncreated.\n- create_optimizer(cfg, trainer): Creates and returns an optimizer for training.\n- create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and\nreturns a learning rate scheduler.\n- add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before\ntraining.\n- add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after\ntraining.\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nadd_callbacks_post_trainer\nAdds callbacks to the trainer after creating the trainer. This is useful for\n\n\nadd_callbacks_pre_trainer\nSet up callbacks before creating the trainer.\n\n\ncreate_lr_scheduler\nCreates and returns a learning rate scheduler.\n\n\ncreate_optimizer\nCreates and returns an optimizer for training.\n\n\nget_input_args\nReturns a pydantic model for the plugin’s input arguments.\n\n\nget_trainer_cls\nReturns a custom class for the trainer.\n\n\nload_datasets\nLoads and preprocesses the dataset for training.\n\n\npost_lora_load\nPerforms actions after LoRA weights are loaded.\n\n\npost_model_build\nPerforms actions after the model is built/loaded, but before any adapters are applied.\n\n\npost_model_load\nPerforms actions after the model is loaded.\n\n\npost_train\nPerforms actions after training is complete.\n\n\npost_train_unload\nPerforms actions after training is complete and the model is unloaded.\n\n\npost_trainer_create\nPerforms actions after the trainer is created.\n\n\npre_lora_load\nPerforms actions before LoRA weights are loaded.\n\n\npre_model_load\nPerforms actions before the model is loaded.\n\n\nregister\nRegisters the plugin with the given configuration.\n\n\n\n\n\nintegrations.base.BasePlugin.add_callbacks_post_trainer(cfg, trainer)\nAdds callbacks to the trainer after creating the trainer. This is useful for\ncallbacks that require access to the model or trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.add_callbacks_pre_trainer(cfg, model)\nSet up callbacks before creating the trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added to the TrainingArgs.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.create_lr_scheduler(\n    cfg,\n    trainer,\n    optimizer,\n    num_training_steps,\n)\nCreates and returns a learning rate scheduler.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\noptimizer\nOptimizer\nThe optimizer for training.\nrequired\n\n\nnum_training_steps\nint\nTotal number of training steps\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nLRScheduler | None\nThe created learning rate scheduler.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.create_optimizer(cfg, trainer)\nCreates and returns an optimizer for training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nOptimizer | None\nThe created optimizer.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.get_input_args()\nReturns a pydantic model for the plugin’s input arguments.\n\n\n\nintegrations.base.BasePlugin.get_trainer_cls(cfg)\nReturns a custom class for the trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe global axolotl configuration.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainer | None\nThe first non-None trainer class returned by a plugin.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.load_datasets(cfg, preprocess=False)\nLoads and preprocesses the dataset for training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\npreprocess\nbool\nWhether this is the preprocess step of the datasets.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\ndataset_meta\nUnion['TrainDatasetMeta', None]\nThe metadata for the training dataset.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_lora_load(cfg, model)\nPerforms actions after LoRA weights are loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_model_build(cfg, model)\nPerforms actions after the model is built/loaded, but before any adapters are applied.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_model_load(cfg, model)\nPerforms actions after the model is loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_train(cfg, model)\nPerforms actions after training is complete.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe axolotl configuration.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_train_unload(cfg)\nPerforms actions after training is complete and the model is unloaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_trainer_create(cfg, trainer)\nPerforms actions after the trainer is created.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.pre_lora_load(cfg, model)\nPerforms actions before LoRA weights are loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.pre_model_load(cfg)\nPerforms actions before the model is loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.register(cfg)\nRegisters the plugin with the given configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\n\n\nintegrations.base.PluginManager()\nThe PluginManager class is responsible for loading and managing plugins. It\nshould be a singleton so it can be accessed from anywhere in the codebase.\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nplugins\nOrderedDict[str, BasePlugin]\nA list of loaded plugins.\n\n\n\n\n\n\nKey methods include:\n- get_instance(): Static method to get the singleton instance of PluginManager.\n- register(plugin_name: str): Registers a new plugin by its name.\n- pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nadd_callbacks_post_trainer\nCalls the add_callbacks_post_trainer method of all registered plugins.\n\n\nadd_callbacks_pre_trainer\nCalls the add_callbacks_pre_trainer method of all registered plugins.\n\n\ncreate_lr_scheduler\nCalls the create_lr_scheduler method of all registered plugins and returns\n\n\ncreate_optimizer\nCalls the create_optimizer method of all registered plugins and returns\n\n\nget_input_args\nReturns a list of Pydantic classes for all registered plugins’ input arguments.’\n\n\nget_instance\nReturns the singleton instance of PluginManager. If the instance doesn’t\n\n\nget_trainer_cls\nCalls the get_trainer_cls method of all registered plugins and returns the\n\n\nload_datasets\nCalls the load_datasets method of each registered plugin.\n\n\npost_lora_load\nCalls the post_lora_load method of all registered plugins.\n\n\npost_model_build\nCalls the post_model_build method of all registered plugins after the\n\n\npost_model_load\nCalls the post_model_load method of all registered plugins after the model\n\n\npost_train\nCalls the post_train method of all registered plugins.\n\n\npost_train_unload\nCalls the post_train_unload method of all registered plugins.\n\n\npost_trainer_create\nCalls the post_trainer_create method of all registered plugins.\n\n\npre_lora_load\nCalls the pre_lora_load method of all registered plugins.\n\n\npre_model_load\nCalls the pre_model_load method of all registered plugins.\n\n\nregister\nRegisters a new plugin by its name.\n\n\n\n\n\nintegrations.base.PluginManager.add_callbacks_post_trainer(cfg, trainer)\nCalls the add_callbacks_post_trainer method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added to the TrainingArgs.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.add_callbacks_pre_trainer(cfg, model)\nCalls the add_callbacks_pre_trainer method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added to the TrainingArgs.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.create_lr_scheduler(\n    trainer,\n    optimizer,\n    num_training_steps,\n)\nCalls the create_lr_scheduler method of all registered plugins and returns\nthe first non-None scheduler.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\noptimizer\nOptimizer\nThe optimizer for training.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nLRScheduler | None\nThe created learning rate scheduler, or None if not found.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.create_optimizer(trainer)\nCalls the create_optimizer method of all registered plugins and returns\nthe first non-None optimizer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nOptimizer | None\nThe created optimizer, or None if none was found.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.get_input_args()\nReturns a list of Pydantic classes for all registered plugins’ input arguments.’\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[str]\nA list of Pydantic classes for all registered plugins’ input arguments.’\n\n\n\n\n\n\n\nintegrations.base.PluginManager.get_instance()\nReturns the singleton instance of PluginManager. If the instance doesn’t\nexist, it creates a new one.\n\n\n\nintegrations.base.PluginManager.get_trainer_cls(cfg)\nCalls the get_trainer_cls method of all registered plugins and returns the\nfirst non-None trainer class.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainer | None\nThe first non-None trainer class returned by a plugin.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.load_datasets(cfg, preprocess=False)\nCalls the load_datasets method of each registered plugin.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\npreprocess\nbool\nWhether this is preprocess step of the datasets.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nUnion['TrainDatasetMeta', None]\nThe dataset metadata loaded from all registered plugins.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_lora_load(cfg, model)\nCalls the post_lora_load method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_model_build(cfg, model)\nCalls the post_model_build method of all registered plugins after the\nmodel has been built / loaded, but before any adapters have been applied.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_model_load(cfg, model)\nCalls the post_model_load method of all registered plugins after the model\nhas been loaded inclusive of any adapters.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_train(cfg, model)\nCalls the post_train method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_train_unload(cfg)\nCalls the post_train_unload method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_trainer_create(cfg, trainer)\nCalls the post_trainer_create method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.pre_lora_load(cfg, model)\nCalls the pre_lora_load method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.pre_model_load(cfg)\nCalls the pre_model_load method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.register(plugin_name)\nRegisters a new plugin by its name.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nplugin_name\nstr\nThe name of the plugin to be registered.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nImportError\nIf the plugin module cannot be imported.\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nload_plugin\nLoads a plugin based on the given plugin name.\n\n\n\n\n\nintegrations.base.load_plugin(plugin_name)\nLoads a plugin based on the given plugin name.\nThe plugin name should be in the format “module_name.class_name”. This function\nsplits the plugin name into module and class, imports the module, retrieves the\nclass from the module, and creates an instance of the class.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nplugin_name\nstr\nThe name of the plugin to be loaded. The name should be in the format “module_name.class_name”.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nBasePlugin\nAn instance of the loaded plugin.\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nImportError\nIf the plugin module cannot be imported."
  },
  {
    "objectID": "docs/api/integrations.base.html#classes",
    "href": "docs/api/integrations.base.html#classes",
    "title": "integrations.base",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nBaseOptimizerFactory\nBase class for factories to create custom optimizers\n\n\nBasePlugin\nBase class for all plugins. Defines the interface for plugin methods.\n\n\nPluginManager\nThe PluginManager class is responsible for loading and managing plugins. It\n\n\n\n\n\nintegrations.base.BaseOptimizerFactory()\nBase class for factories to create custom optimizers\n\n\n\nintegrations.base.BasePlugin()\nBase class for all plugins. Defines the interface for plugin methods.\nA plugin is a reusable, modular, and self-contained piece of code that extends\nthe functionality of Axolotl. Plugins can be used to integrate third-party models,\nmodify the training process, or add new features.\nTo create a new plugin, you need to inherit from the BasePlugin class and\nimplement the required methods.\n\n\nPlugin methods include:\n- register(cfg): Registers the plugin with the given configuration.\n- load_datasets(cfg): Loads and preprocesses the dataset for training.\n- pre_model_load(cfg): Performs actions before the model is loaded.\n- post_model_build(cfg, model): Performs actions after the model is loaded, but\nbefore LoRA adapters are applied.\n- pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.\n- post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.\n- post_model_load(cfg, model): Performs actions after the model is loaded,\ninclusive of any adapters.\n- post_trainer_create(cfg, trainer): Performs actions after the trainer is\ncreated.\n- create_optimizer(cfg, trainer): Creates and returns an optimizer for training.\n- create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and\nreturns a learning rate scheduler.\n- add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before\ntraining.\n- add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after\ntraining.\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nadd_callbacks_post_trainer\nAdds callbacks to the trainer after creating the trainer. This is useful for\n\n\nadd_callbacks_pre_trainer\nSet up callbacks before creating the trainer.\n\n\ncreate_lr_scheduler\nCreates and returns a learning rate scheduler.\n\n\ncreate_optimizer\nCreates and returns an optimizer for training.\n\n\nget_input_args\nReturns a pydantic model for the plugin’s input arguments.\n\n\nget_trainer_cls\nReturns a custom class for the trainer.\n\n\nload_datasets\nLoads and preprocesses the dataset for training.\n\n\npost_lora_load\nPerforms actions after LoRA weights are loaded.\n\n\npost_model_build\nPerforms actions after the model is built/loaded, but before any adapters are applied.\n\n\npost_model_load\nPerforms actions after the model is loaded.\n\n\npost_train\nPerforms actions after training is complete.\n\n\npost_train_unload\nPerforms actions after training is complete and the model is unloaded.\n\n\npost_trainer_create\nPerforms actions after the trainer is created.\n\n\npre_lora_load\nPerforms actions before LoRA weights are loaded.\n\n\npre_model_load\nPerforms actions before the model is loaded.\n\n\nregister\nRegisters the plugin with the given configuration.\n\n\n\n\n\nintegrations.base.BasePlugin.add_callbacks_post_trainer(cfg, trainer)\nAdds callbacks to the trainer after creating the trainer. This is useful for\ncallbacks that require access to the model or trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.add_callbacks_pre_trainer(cfg, model)\nSet up callbacks before creating the trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added to the TrainingArgs.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.create_lr_scheduler(\n    cfg,\n    trainer,\n    optimizer,\n    num_training_steps,\n)\nCreates and returns a learning rate scheduler.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\noptimizer\nOptimizer\nThe optimizer for training.\nrequired\n\n\nnum_training_steps\nint\nTotal number of training steps\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nLRScheduler | None\nThe created learning rate scheduler.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.create_optimizer(cfg, trainer)\nCreates and returns an optimizer for training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nOptimizer | None\nThe created optimizer.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.get_input_args()\nReturns a pydantic model for the plugin’s input arguments.\n\n\n\nintegrations.base.BasePlugin.get_trainer_cls(cfg)\nReturns a custom class for the trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe global axolotl configuration.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainer | None\nThe first non-None trainer class returned by a plugin.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.load_datasets(cfg, preprocess=False)\nLoads and preprocesses the dataset for training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\npreprocess\nbool\nWhether this is the preprocess step of the datasets.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\ndataset_meta\nUnion['TrainDatasetMeta', None]\nThe metadata for the training dataset.\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_lora_load(cfg, model)\nPerforms actions after LoRA weights are loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_model_build(cfg, model)\nPerforms actions after the model is built/loaded, but before any adapters are applied.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_model_load(cfg, model)\nPerforms actions after the model is loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_train(cfg, model)\nPerforms actions after training is complete.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe axolotl configuration.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_train_unload(cfg)\nPerforms actions after training is complete and the model is unloaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.post_trainer_create(cfg, trainer)\nPerforms actions after the trainer is created.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.pre_lora_load(cfg, model)\nPerforms actions before LoRA weights are loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.pre_model_load(cfg)\nPerforms actions before the model is loaded.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\nintegrations.base.BasePlugin.register(cfg)\nRegisters the plugin with the given configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugin.\nrequired\n\n\n\n\n\n\n\n\n\nintegrations.base.PluginManager()\nThe PluginManager class is responsible for loading and managing plugins. It\nshould be a singleton so it can be accessed from anywhere in the codebase.\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nplugins\nOrderedDict[str, BasePlugin]\nA list of loaded plugins.\n\n\n\n\n\n\nKey methods include:\n- get_instance(): Static method to get the singleton instance of PluginManager.\n- register(plugin_name: str): Registers a new plugin by its name.\n- pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nadd_callbacks_post_trainer\nCalls the add_callbacks_post_trainer method of all registered plugins.\n\n\nadd_callbacks_pre_trainer\nCalls the add_callbacks_pre_trainer method of all registered plugins.\n\n\ncreate_lr_scheduler\nCalls the create_lr_scheduler method of all registered plugins and returns\n\n\ncreate_optimizer\nCalls the create_optimizer method of all registered plugins and returns\n\n\nget_input_args\nReturns a list of Pydantic classes for all registered plugins’ input arguments.’\n\n\nget_instance\nReturns the singleton instance of PluginManager. If the instance doesn’t\n\n\nget_trainer_cls\nCalls the get_trainer_cls method of all registered plugins and returns the\n\n\nload_datasets\nCalls the load_datasets method of each registered plugin.\n\n\npost_lora_load\nCalls the post_lora_load method of all registered plugins.\n\n\npost_model_build\nCalls the post_model_build method of all registered plugins after the\n\n\npost_model_load\nCalls the post_model_load method of all registered plugins after the model\n\n\npost_train\nCalls the post_train method of all registered plugins.\n\n\npost_train_unload\nCalls the post_train_unload method of all registered plugins.\n\n\npost_trainer_create\nCalls the post_trainer_create method of all registered plugins.\n\n\npre_lora_load\nCalls the pre_lora_load method of all registered plugins.\n\n\npre_model_load\nCalls the pre_model_load method of all registered plugins.\n\n\nregister\nRegisters a new plugin by its name.\n\n\n\n\n\nintegrations.base.PluginManager.add_callbacks_post_trainer(cfg, trainer)\nCalls the add_callbacks_post_trainer method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added to the TrainingArgs.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.add_callbacks_pre_trainer(cfg, model)\nCalls the add_callbacks_pre_trainer method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[Callable]\nA list of callback functions to be added to the TrainingArgs.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.create_lr_scheduler(\n    trainer,\n    optimizer,\n    num_training_steps,\n)\nCalls the create_lr_scheduler method of all registered plugins and returns\nthe first non-None scheduler.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\noptimizer\nOptimizer\nThe optimizer for training.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nLRScheduler | None\nThe created learning rate scheduler, or None if not found.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.create_optimizer(trainer)\nCalls the create_optimizer method of all registered plugins and returns\nthe first non-None optimizer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nOptimizer | None\nThe created optimizer, or None if none was found.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.get_input_args()\nReturns a list of Pydantic classes for all registered plugins’ input arguments.’\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[str]\nA list of Pydantic classes for all registered plugins’ input arguments.’\n\n\n\n\n\n\n\nintegrations.base.PluginManager.get_instance()\nReturns the singleton instance of PluginManager. If the instance doesn’t\nexist, it creates a new one.\n\n\n\nintegrations.base.PluginManager.get_trainer_cls(cfg)\nCalls the get_trainer_cls method of all registered plugins and returns the\nfirst non-None trainer class.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTrainer | None\nThe first non-None trainer class returned by a plugin.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.load_datasets(cfg, preprocess=False)\nCalls the load_datasets method of each registered plugin.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\npreprocess\nbool\nWhether this is preprocess step of the datasets.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nUnion['TrainDatasetMeta', None]\nThe dataset metadata loaded from all registered plugins.\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_lora_load(cfg, model)\nCalls the post_lora_load method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_model_build(cfg, model)\nCalls the post_model_build method of all registered plugins after the\nmodel has been built / loaded, but before any adapters have been applied.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_model_load(cfg, model)\nCalls the post_model_load method of all registered plugins after the model\nhas been loaded inclusive of any adapters.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_train(cfg, model)\nCalls the post_train method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel | PeftModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_train_unload(cfg)\nCalls the post_train_unload method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.post_trainer_create(cfg, trainer)\nCalls the post_trainer_create method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\ntrainer\nTrainer\nThe trainer object for training.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.pre_lora_load(cfg, model)\nCalls the pre_lora_load method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\nmodel\nPreTrainedModel\nThe loaded model.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.pre_model_load(cfg)\nCalls the pre_model_load method of all registered plugins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nThe configuration for the plugins.\nrequired\n\n\n\n\n\n\n\nintegrations.base.PluginManager.register(plugin_name)\nRegisters a new plugin by its name.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nplugin_name\nstr\nThe name of the plugin to be registered.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nImportError\nIf the plugin module cannot be imported."
  },
  {
    "objectID": "docs/api/integrations.base.html#functions",
    "href": "docs/api/integrations.base.html#functions",
    "title": "integrations.base",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nload_plugin\nLoads a plugin based on the given plugin name.\n\n\n\n\n\nintegrations.base.load_plugin(plugin_name)\nLoads a plugin based on the given plugin name.\nThe plugin name should be in the format “module_name.class_name”. This function\nsplits the plugin name into module and class, imports the module, retrieves the\nclass from the module, and creates an instance of the class.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nplugin_name\nstr\nThe name of the plugin to be loaded. The name should be in the format “module_name.class_name”.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nBasePlugin\nAn instance of the loaded plugin.\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nImportError\nIf the plugin module cannot be imported."
  },
  {
    "objectID": "docs/api/utils.samplers.multipack.html",
    "href": "docs/api/utils.samplers.multipack.html",
    "title": "utils.samplers.multipack",
    "section": "",
    "text": "utils.samplers.multipack\nMultipack Batch Sampler - An efficient batch sampler for packing variable-length sequences\ninto fixed-capacity batches to optimize memory usage and training throughput.\n\n\n\n\n\nName\nDescription\n\n\n\n\nMultipackBatchSampler\nBatch sampler class for efficient packing of variable-length sequences\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler(\n    sampler,\n    batch_size,\n    batch_max_len,\n    lengths,\n    packing_efficiency_estimate=1.0,\n    drop_last=False,\n    num_count_samples=16,\n    sequential=False,\n    group_size=100000,\n    bin_size=200,\n    num_processes=None,\n    safe_mode=True,\n    **kwargs,\n)\nBatch sampler class for efficient packing of variable-length sequences\nThis sampler packs sequences into fixed-capacity bins (batches) to maximize\nGPU memory utilization and training throughput by reducing padding.\nIt supports both parallel packing (using FFD algorithm) and\nsequential packing (preserving original sequence order).\n\n\n\n\n\nName\nDescription\n\n\n\n\nefficiency\nCalculate the packing efficiency (ratio of tokens used to total token slots).\n\n\ngather_efficiency\nGather and synchronize packing efficiency estimates across all distributed\n\n\ngather_len_batches\nGather and synchronize batch counts across all distributed ranks. Returns\n\n\ngenerate_batches\nGenerate packed batches for training.\n\n\nset_epoch\nSet the epoch number, used for reproducible shuffling across epochs\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.efficiency()\nCalculate the packing efficiency (ratio of tokens used to total token slots).\nHigher is better - 1.0 would mean perfect packing with no wasted space.\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.gather_efficiency()\nGather and synchronize packing efficiency estimates across all distributed\nranks.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nfloat\nA conservative efficiency estimate based on the measurements.\n\n\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.gather_len_batches(num)\nGather and synchronize batch counts across all distributed ranks. Returns\nthe minimum number of batches available on any rank.\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.generate_batches(set_stats=False)\nGenerate packed batches for training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nset_stats\nbool\nWhether to update efficiency statistics.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[list[list[int]]]\nList of batches, where each batch contains multiple bins, and each bin contains multiple sequence indices.\n\n\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.set_epoch(epoch)\nSet the epoch number, used for reproducible shuffling across epochs\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nallocate_sequentially\nSequential allocator that preserves example order.\n\n\nffd_check\nFirst-fit-decreasing bin packing algorithm check.\n\n\npack_group\nPack a group of sequences into bins using First-Fit Decreasing algorithm.\n\n\npack_parallel\nPack sequences into bins using parallel processing.\n\n\n\n\n\nutils.samplers.multipack.allocate_sequentially(\n    sequence_lengths,\n    rank,\n    bin_capacity,\n    num_ranks,\n)\nSequential allocator that preserves example order.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nThe lengths of all examples.\nrequired\n\n\nrank\nint\nThe current rank (for distributed training).\nrequired\n\n\nbin_capacity\nint\nThe capacity of each bin (maximum sequence length).\nrequired\n\n\nnum_ranks\nint\nNumber of ranks (processes / GPUs).\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nrank_batches\nlist[list[int]]\nList of batches for the current rank.\n\n\ntotal_tokens_used\nint\nNumber of actual example tokens.\n\n\ntotal_token_slots\nint\nMaximum theoretical number of example tokens (number of bins * bin capacity).\n\n\n\n\n\n\n\nutils.samplers.multipack.ffd_check(sequence_lengths, bin_capacity, num_bins)\nFirst-fit-decreasing bin packing algorithm check.\nChecks if sequences with the given lengths could fit in the specified number of\nbins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nArray of sequence lengths.\nrequired\n\n\nbin_capacity\nint\nMaximum capacity of each bin.\nrequired\n\n\nnum_bins\nint\nNumber of bins available.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nbool\nTrue if all sequences can be packed, False otherwise.\n\n\n\n\n\n\n\nutils.samplers.multipack.pack_group(\n    sequence_lengths,\n    group_offset,\n    bin_capacity,\n    max_bins,\n    bin_size,\n    safe_mode=True,\n)\nPack a group of sequences into bins using First-Fit Decreasing algorithm.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nArray of sequence lengths.\nrequired\n\n\ngroup_offset\nint\nOffset to apply to indices when returning results.\nrequired\n\n\nbin_capacity\nint\nMaximum capacity of each bin.\nrequired\n\n\nmax_bins\nint\nMaximum number of bins to use.\nrequired\n\n\nbin_size\nint\nMaximum number of sequences per bin.\nrequired\n\n\nsafe_mode\nbool\nIf True, use a more conservative packing approach.\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[list[int]]\nList of bins, where each bin contains indices of sequences assigned to it.\n\n\n\n\n\n\n\nutils.samplers.multipack.pack_parallel(\n    sequence_lengths,\n    bin_capacity,\n    group_size,\n    bin_size,\n    num_processes=None,\n    safe_mode=True,\n    mp_start_method='spawn',\n)\nPack sequences into bins using parallel processing.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nArray of sequence lengths.\nrequired\n\n\nbin_capacity\nint\nMaximum capacity of each bin as total number of tokens.\nrequired\n\n\ngroup_size\nint\nNumber of sequences to process in each group.\nrequired\n\n\nbin_size\nint\nMaximum number of bins to use.\nrequired\n\n\nnum_processes\nint | None\nNumber of parallel processes to use.\nNone\n\n\nsafe_mode\nbool\nIf True, use a more conservative packing approach.\nTrue\n\n\nmp_start_method\nstr | None\nMultiprocessing start method (‘fork’, ‘spawn’, ‘forkserver’). ‘spawn’ is often safer with Numba/PyTorch. Set to None to use system default.\n'spawn'\n\n\n\nReturns:\nList of bins, where each bin contains indices of sequences assigned to it."
  },
  {
    "objectID": "docs/api/utils.samplers.multipack.html#classes",
    "href": "docs/api/utils.samplers.multipack.html#classes",
    "title": "utils.samplers.multipack",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMultipackBatchSampler\nBatch sampler class for efficient packing of variable-length sequences\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler(\n    sampler,\n    batch_size,\n    batch_max_len,\n    lengths,\n    packing_efficiency_estimate=1.0,\n    drop_last=False,\n    num_count_samples=16,\n    sequential=False,\n    group_size=100000,\n    bin_size=200,\n    num_processes=None,\n    safe_mode=True,\n    **kwargs,\n)\nBatch sampler class for efficient packing of variable-length sequences\nThis sampler packs sequences into fixed-capacity bins (batches) to maximize\nGPU memory utilization and training throughput by reducing padding.\nIt supports both parallel packing (using FFD algorithm) and\nsequential packing (preserving original sequence order).\n\n\n\n\n\nName\nDescription\n\n\n\n\nefficiency\nCalculate the packing efficiency (ratio of tokens used to total token slots).\n\n\ngather_efficiency\nGather and synchronize packing efficiency estimates across all distributed\n\n\ngather_len_batches\nGather and synchronize batch counts across all distributed ranks. Returns\n\n\ngenerate_batches\nGenerate packed batches for training.\n\n\nset_epoch\nSet the epoch number, used for reproducible shuffling across epochs\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.efficiency()\nCalculate the packing efficiency (ratio of tokens used to total token slots).\nHigher is better - 1.0 would mean perfect packing with no wasted space.\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.gather_efficiency()\nGather and synchronize packing efficiency estimates across all distributed\nranks.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nfloat\nA conservative efficiency estimate based on the measurements.\n\n\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.gather_len_batches(num)\nGather and synchronize batch counts across all distributed ranks. Returns\nthe minimum number of batches available on any rank.\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.generate_batches(set_stats=False)\nGenerate packed batches for training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nset_stats\nbool\nWhether to update efficiency statistics.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[list[list[int]]]\nList of batches, where each batch contains multiple bins, and each bin contains multiple sequence indices.\n\n\n\n\n\n\n\nutils.samplers.multipack.MultipackBatchSampler.set_epoch(epoch)\nSet the epoch number, used for reproducible shuffling across epochs"
  },
  {
    "objectID": "docs/api/utils.samplers.multipack.html#functions",
    "href": "docs/api/utils.samplers.multipack.html#functions",
    "title": "utils.samplers.multipack",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nallocate_sequentially\nSequential allocator that preserves example order.\n\n\nffd_check\nFirst-fit-decreasing bin packing algorithm check.\n\n\npack_group\nPack a group of sequences into bins using First-Fit Decreasing algorithm.\n\n\npack_parallel\nPack sequences into bins using parallel processing.\n\n\n\n\n\nutils.samplers.multipack.allocate_sequentially(\n    sequence_lengths,\n    rank,\n    bin_capacity,\n    num_ranks,\n)\nSequential allocator that preserves example order.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nThe lengths of all examples.\nrequired\n\n\nrank\nint\nThe current rank (for distributed training).\nrequired\n\n\nbin_capacity\nint\nThe capacity of each bin (maximum sequence length).\nrequired\n\n\nnum_ranks\nint\nNumber of ranks (processes / GPUs).\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nrank_batches\nlist[list[int]]\nList of batches for the current rank.\n\n\ntotal_tokens_used\nint\nNumber of actual example tokens.\n\n\ntotal_token_slots\nint\nMaximum theoretical number of example tokens (number of bins * bin capacity).\n\n\n\n\n\n\n\nutils.samplers.multipack.ffd_check(sequence_lengths, bin_capacity, num_bins)\nFirst-fit-decreasing bin packing algorithm check.\nChecks if sequences with the given lengths could fit in the specified number of\nbins.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nArray of sequence lengths.\nrequired\n\n\nbin_capacity\nint\nMaximum capacity of each bin.\nrequired\n\n\nnum_bins\nint\nNumber of bins available.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nbool\nTrue if all sequences can be packed, False otherwise.\n\n\n\n\n\n\n\nutils.samplers.multipack.pack_group(\n    sequence_lengths,\n    group_offset,\n    bin_capacity,\n    max_bins,\n    bin_size,\n    safe_mode=True,\n)\nPack a group of sequences into bins using First-Fit Decreasing algorithm.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nArray of sequence lengths.\nrequired\n\n\ngroup_offset\nint\nOffset to apply to indices when returning results.\nrequired\n\n\nbin_capacity\nint\nMaximum capacity of each bin.\nrequired\n\n\nmax_bins\nint\nMaximum number of bins to use.\nrequired\n\n\nbin_size\nint\nMaximum number of sequences per bin.\nrequired\n\n\nsafe_mode\nbool\nIf True, use a more conservative packing approach.\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[list[int]]\nList of bins, where each bin contains indices of sequences assigned to it.\n\n\n\n\n\n\n\nutils.samplers.multipack.pack_parallel(\n    sequence_lengths,\n    bin_capacity,\n    group_size,\n    bin_size,\n    num_processes=None,\n    safe_mode=True,\n    mp_start_method='spawn',\n)\nPack sequences into bins using parallel processing.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nsequence_lengths\nnp.ndarray\nArray of sequence lengths.\nrequired\n\n\nbin_capacity\nint\nMaximum capacity of each bin as total number of tokens.\nrequired\n\n\ngroup_size\nint\nNumber of sequences to process in each group.\nrequired\n\n\nbin_size\nint\nMaximum number of bins to use.\nrequired\n\n\nnum_processes\nint | None\nNumber of parallel processes to use.\nNone\n\n\nsafe_mode\nbool\nIf True, use a more conservative packing approach.\nTrue\n\n\nmp_start_method\nstr | None\nMultiprocessing start method (‘fork’, ‘spawn’, ‘forkserver’). ‘spawn’ is often safer with Numba/PyTorch. Set to None to use system default.\n'spawn'\n\n\n\nReturns:\nList of bins, where each bin contains indices of sequences assigned to it."
  },
  {
    "objectID": "docs/api/prompt_strategies.completion.html",
    "href": "docs/api/prompt_strategies.completion.html",
    "title": "prompt_strategies.completion",
    "section": "",
    "text": "prompt_strategies.completion\nBasic completion text\n\n\n\n\n\nName\nDescription\n\n\n\n\nCompletionPromptTokenizingStrategy\nTokenizing strategy for Completion prompts.\n\n\nCompletionPrompter\nPrompter for completion\n\n\n\n\n\nprompt_strategies.completion.CompletionPromptTokenizingStrategy(\n    *args,\n    max_length=None,\n    **kwargs,\n)\nTokenizing strategy for Completion prompts.\n\n\n\nprompt_strategies.completion.CompletionPrompter()\nPrompter for completion"
  },
  {
    "objectID": "docs/api/prompt_strategies.completion.html#classes",
    "href": "docs/api/prompt_strategies.completion.html#classes",
    "title": "prompt_strategies.completion",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nCompletionPromptTokenizingStrategy\nTokenizing strategy for Completion prompts.\n\n\nCompletionPrompter\nPrompter for completion\n\n\n\n\n\nprompt_strategies.completion.CompletionPromptTokenizingStrategy(\n    *args,\n    max_length=None,\n    **kwargs,\n)\nTokenizing strategy for Completion prompts.\n\n\n\nprompt_strategies.completion.CompletionPrompter()\nPrompter for completion"
  },
  {
    "objectID": "docs/api/utils.collators.mamba.html",
    "href": "docs/api/utils.collators.mamba.html",
    "title": "utils.collators.mamba",
    "section": "",
    "text": "utils.collators.mamba\ncollators for Mamba\n\n\n\n\n\nName\nDescription\n\n\n\n\nMambaDataCollator\nCollator for State Space Models (Mamba)\n\n\n\n\n\nutils.collators.mamba.MambaDataCollator(tokenizer)\nCollator for State Space Models (Mamba)"
  },
  {
    "objectID": "docs/api/utils.collators.mamba.html#classes",
    "href": "docs/api/utils.collators.mamba.html#classes",
    "title": "utils.collators.mamba",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMambaDataCollator\nCollator for State Space Models (Mamba)\n\n\n\n\n\nutils.collators.mamba.MambaDataCollator(tokenizer)\nCollator for State Space Models (Mamba)"
  },
  {
    "objectID": "docs/api/utils.schemas.model.html",
    "href": "docs/api/utils.schemas.model.html",
    "title": "utils.schemas.model",
    "section": "",
    "text": "utils.schemas.model\nPydantic models for model input / output, etc. configuration\n\n\n\n\n\nName\nDescription\n\n\n\n\nModelInputConfig\nModel configuration subset\n\n\nModelOutputConfig\nmodel save configuration subset\n\n\nSpecialTokensConfig\nSpecial tokens configuration subset\n\n\n\n\n\nutils.schemas.model.ModelInputConfig()\nModel configuration subset\n\n\n\nutils.schemas.model.ModelOutputConfig()\nmodel save configuration subset\n\n\n\nutils.schemas.model.SpecialTokensConfig()\nSpecial tokens configuration subset"
  },
  {
    "objectID": "docs/api/utils.schemas.model.html#classes",
    "href": "docs/api/utils.schemas.model.html#classes",
    "title": "utils.schemas.model",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nModelInputConfig\nModel configuration subset\n\n\nModelOutputConfig\nmodel save configuration subset\n\n\nSpecialTokensConfig\nSpecial tokens configuration subset\n\n\n\n\n\nutils.schemas.model.ModelInputConfig()\nModel configuration subset\n\n\n\nutils.schemas.model.ModelOutputConfig()\nmodel save configuration subset\n\n\n\nutils.schemas.model.SpecialTokensConfig()\nSpecial tokens configuration subset"
  },
  {
    "objectID": "docs/api/utils.schemas.training.html",
    "href": "docs/api/utils.schemas.training.html",
    "title": "utils.schemas.training",
    "section": "",
    "text": "utils.schemas.training\nPydantic models for training hyperparameters\n\n\n\n\n\nName\nDescription\n\n\n\n\nHyperparametersConfig\nTraining hyperparams configuration subset\n\n\nLrGroup\nCustom learning rate group configuration\n\n\n\n\n\nutils.schemas.training.HyperparametersConfig()\nTraining hyperparams configuration subset\n\n\n\nutils.schemas.training.LrGroup()\nCustom learning rate group configuration"
  },
  {
    "objectID": "docs/api/utils.schemas.training.html#classes",
    "href": "docs/api/utils.schemas.training.html#classes",
    "title": "utils.schemas.training",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nHyperparametersConfig\nTraining hyperparams configuration subset\n\n\nLrGroup\nCustom learning rate group configuration\n\n\n\n\n\nutils.schemas.training.HyperparametersConfig()\nTraining hyperparams configuration subset\n\n\n\nutils.schemas.training.LrGroup()\nCustom learning rate group configuration"
  },
  {
    "objectID": "docs/api/prompt_strategies.kto.llama3.html",
    "href": "docs/api/prompt_strategies.kto.llama3.html",
    "title": "prompt_strategies.kto.llama3",
    "section": "",
    "text": "prompt_strategies.kto.llama3\nKTO strategies for llama-3 chat template\n\n\n\n\n\nName\nDescription\n\n\n\n\nargilla_chat\nfor argilla/kto-mix-15k conversations\n\n\nintel\nFor Intel Orca KTO\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.kto.llama3.argilla_chat(cfg, **kwargs)\nfor argilla/kto-mix-15k conversations\n\n\n\nprompt_strategies.kto.llama3.intel(cfg, **kwargs)\nFor Intel Orca KTO\nex: argilla/distilabel-intel-orca-kto\n\n\n\nprompt_strategies.kto.llama3.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations\nex: argilla/ultrafeedback-binarized-preferences-cleaned-kto"
  },
  {
    "objectID": "docs/api/prompt_strategies.kto.llama3.html#functions",
    "href": "docs/api/prompt_strategies.kto.llama3.html#functions",
    "title": "prompt_strategies.kto.llama3",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nargilla_chat\nfor argilla/kto-mix-15k conversations\n\n\nintel\nFor Intel Orca KTO\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.kto.llama3.argilla_chat(cfg, **kwargs)\nfor argilla/kto-mix-15k conversations\n\n\n\nprompt_strategies.kto.llama3.intel(cfg, **kwargs)\nFor Intel Orca KTO\nex: argilla/distilabel-intel-orca-kto\n\n\n\nprompt_strategies.kto.llama3.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations\nex: argilla/ultrafeedback-binarized-preferences-cleaned-kto"
  },
  {
    "objectID": "docs/api/core.trainers.utils.html",
    "href": "docs/api/core.trainers.utils.html",
    "title": "core.trainers.utils",
    "section": "",
    "text": "core.trainers.utils\ncore.trainers.utils\nUtils for Axolotl trainers"
  },
  {
    "objectID": "docs/api/core.trainers.mixins.optimizer.html",
    "href": "docs/api/core.trainers.mixins.optimizer.html",
    "title": "core.trainers.mixins.optimizer",
    "section": "",
    "text": "core.trainers.mixins.optimizer\nModule for Axolotl trainer optimizer mixin\n\n\n\n\n\nName\nDescription\n\n\n\n\nOptimizerInitMixin\nMixin to handle common optimizer initialization logic for Trainers (mostly TRL) that do not\n\n\nOptimizerMixin\nMixin class for shared handling of building custom optimizers\n\n\n\n\n\ncore.trainers.mixins.optimizer.OptimizerInitMixin(*args, **kwargs)\nMixin to handle common optimizer initialization logic for Trainers (mostly TRL) that do not\naccept optimizer_cls_and_kwargs as kwarg in constructor.\n\n\n\ncore.trainers.mixins.optimizer.OptimizerMixin()\nMixin class for shared handling of building custom optimizers"
  },
  {
    "objectID": "docs/api/core.trainers.mixins.optimizer.html#classes",
    "href": "docs/api/core.trainers.mixins.optimizer.html#classes",
    "title": "core.trainers.mixins.optimizer",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nOptimizerInitMixin\nMixin to handle common optimizer initialization logic for Trainers (mostly TRL) that do not\n\n\nOptimizerMixin\nMixin class for shared handling of building custom optimizers\n\n\n\n\n\ncore.trainers.mixins.optimizer.OptimizerInitMixin(*args, **kwargs)\nMixin to handle common optimizer initialization logic for Trainers (mostly TRL) that do not\naccept optimizer_cls_and_kwargs as kwarg in constructor.\n\n\n\ncore.trainers.mixins.optimizer.OptimizerMixin()\nMixin class for shared handling of building custom optimizers"
  },
  {
    "objectID": "docs/api/utils.freeze.html",
    "href": "docs/api/utils.freeze.html",
    "title": "utils.freeze",
    "section": "",
    "text": "utils.freeze\nmodule to freeze/unfreeze parameters by name\n\n\n\n\n\nName\nDescription\n\n\n\n\nLayerNamePattern\nRepresents a regex pattern for layer names, potentially including a parameter index range.\n\n\n\n\n\nutils.freeze.LayerNamePattern(pattern)\nRepresents a regex pattern for layer names, potentially including a parameter index range.\n\n\n\n\n\nName\nDescription\n\n\n\n\nmatch\nChecks if the given layer name matches the regex pattern.\n\n\n\n\n\nutils.freeze.LayerNamePattern.match(name)\nChecks if the given layer name matches the regex pattern.\nParameters:\n- name (str): The layer name to check.\nReturns:\n- bool: True if the layer name matches the pattern, False otherwise.\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nfreeze_layers_except\nFreezes all layers of the given model except for the layers that match given regex patterns.\n\n\n\n\n\nutils.freeze.freeze_layers_except(model, regex_patterns)\nFreezes all layers of the given model except for the layers that match given regex patterns.\nPeriods in the patterns are treated as literal periods, not as wildcard characters.\nParameters:\n- model (nn.Module): The PyTorch model to be modified.\n- regex_patterns (list of str): List of regex patterns to match layer names to keep unfrozen.\nNote that you cannot use a dot as a wildcard character in the patterns since it is reserved for separating layer names.\nAlso, to match the entire layer name, the pattern should start with “^” and end with “\\(\", otherwise it will match any part of the layer name.\n  The range pattern part is optional and it is not compiled as a regex pattern which means you must put \"\\)” before the range pattern if you want to match the entire layer name.\nE.g., [“^model.embed_tokens.weight\\([:32000]\", \"layers.2[0-9]+.block_sparse_moe.gate.[a-z]+\\)”]\nReturns:\nNone; the model is modified in place."
  },
  {
    "objectID": "docs/api/utils.freeze.html#classes",
    "href": "docs/api/utils.freeze.html#classes",
    "title": "utils.freeze",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nLayerNamePattern\nRepresents a regex pattern for layer names, potentially including a parameter index range.\n\n\n\n\n\nutils.freeze.LayerNamePattern(pattern)\nRepresents a regex pattern for layer names, potentially including a parameter index range.\n\n\n\n\n\nName\nDescription\n\n\n\n\nmatch\nChecks if the given layer name matches the regex pattern.\n\n\n\n\n\nutils.freeze.LayerNamePattern.match(name)\nChecks if the given layer name matches the regex pattern.\nParameters:\n- name (str): The layer name to check.\nReturns:\n- bool: True if the layer name matches the pattern, False otherwise."
  },
  {
    "objectID": "docs/api/utils.freeze.html#functions",
    "href": "docs/api/utils.freeze.html#functions",
    "title": "utils.freeze",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nfreeze_layers_except\nFreezes all layers of the given model except for the layers that match given regex patterns.\n\n\n\n\n\nutils.freeze.freeze_layers_except(model, regex_patterns)\nFreezes all layers of the given model except for the layers that match given regex patterns.\nPeriods in the patterns are treated as literal periods, not as wildcard characters.\nParameters:\n- model (nn.Module): The PyTorch model to be modified.\n- regex_patterns (list of str): List of regex patterns to match layer names to keep unfrozen.\nNote that you cannot use a dot as a wildcard character in the patterns since it is reserved for separating layer names.\nAlso, to match the entire layer name, the pattern should start with “^” and end with “\\(\", otherwise it will match any part of the layer name.\n  The range pattern part is optional and it is not compiled as a regex pattern which means you must put \"\\)” before the range pattern if you want to match the entire layer name.\nE.g., [“^model.embed_tokens.weight\\([:32000]\", \"layers.2[0-9]+.block_sparse_moe.gate.[a-z]+\\)”]\nReturns:\nNone; the model is modified in place."
  },
  {
    "objectID": "docs/api/cli.utils.html",
    "href": "docs/api/cli.utils.html",
    "title": "cli.utils",
    "section": "",
    "text": "cli.utils\nUtility methods for axolotl CLI.\n\n\n\n\n\nName\nDescription\n\n\n\n\nadd_options_from_config\nCreate Click options from the fields of a Pydantic model.\n\n\nadd_options_from_dataclass\nCreate Click options from the fields of a dataclass.\n\n\nbuild_command\nBuild command list from base command and options.\n\n\ndownload_file\nDownload a single file and return its processing status.\n\n\nfetch_from_github\nSync files from a specific directory in the GitHub repository.\n\n\nfilter_none_kwargs\nWraps function to remove None-valued kwargs.\n\n\nload_model_and_tokenizer\nHelper function for loading a model, tokenizer, and processor specified in the given axolotl\n\n\nstrip_optional_type\nExtracts the non-None type from an Optional / Union type.\n\n\n\n\n\ncli.utils.add_options_from_config(config_class)\nCreate Click options from the fields of a Pydantic model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig_class\nType[BaseModel]\nPyDantic model with fields to parse from the CLI\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nCallable\nFunction decorator for Axolotl CLI command.\n\n\n\n\n\n\n\ncli.utils.add_options_from_dataclass(config_class)\nCreate Click options from the fields of a dataclass.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig_class\nType[Any]\nDataclass with fields to parse from the CLI.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nCallable\nFunction decorator for Axolotl CLI command.\n\n\n\n\n\n\n\ncli.utils.build_command(base_cmd, options)\nBuild command list from base command and options.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbase_cmd\nlist[str]\nCommand without options.\nrequired\n\n\noptions\ndict[str, Any]\nOptions to parse and append to base command.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[str]\nList of strings giving shell command.\n\n\n\n\n\n\n\ncli.utils.download_file(file_info, raw_base_url, dest_path, dir_prefix)\nDownload a single file and return its processing status.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nfile_info\ntuple\nTuple of (file_path, remote_sha).\nrequired\n\n\nraw_base_url\nstr\nBase URL for raw GitHub content.\nrequired\n\n\ndest_path\nPath\nLocal destination directory.\nrequired\n\n\ndir_prefix\nstr\nDirectory prefix to filter files.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[str, str]\nTuple of (file_path, status) where status is ‘new’, ‘updated’, or ‘unchanged’.\n\n\n\n\n\n\n\ncli.utils.fetch_from_github(dir_prefix, dest_dir=None, max_workers=5)\nSync files from a specific directory in the GitHub repository.\nOnly downloads files that don’t exist locally or have changed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndir_prefix\nstr\nDirectory prefix to filter files (e.g., ‘examples/’, ‘deepspeed_configs/’).\nrequired\n\n\ndest_dir\nstr | None\nLocal destination directory.\nNone\n\n\nmax_workers\nint\nMaximum number of concurrent downloads.\n5\n\n\n\n\n\n\n\ncli.utils.filter_none_kwargs(func)\nWraps function to remove None-valued kwargs.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nfunc\nCallable\nFunction to wrap.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nCallable\nWrapped function.\n\n\n\n\n\n\n\ncli.utils.load_model_and_tokenizer(cfg, inference=False)\nHelper function for loading a model, tokenizer, and processor specified in the given axolotl\nconfig.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ninference\nbool\nBoolean denoting inference mode.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PreTrainedModel, PreTrainedTokenizer | PreTrainedTokenizerFast | Any, ProcessorMixin | None]\nTuple of (PreTrainedModel, PreTrainedTokenizer, ProcessorMixin).\n\n\n\n\n\n\n\ncli.utils.strip_optional_type(field_type)\nExtracts the non-None type from an Optional / Union type.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nfield_type\ntype | str | None\nType of field for Axolotl CLI command.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\n\nIf the input type is Union[T, None] or Optional[T], returns T. Otherwise returns the input type unchanged."
  },
  {
    "objectID": "docs/api/cli.utils.html#functions",
    "href": "docs/api/cli.utils.html#functions",
    "title": "cli.utils",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nadd_options_from_config\nCreate Click options from the fields of a Pydantic model.\n\n\nadd_options_from_dataclass\nCreate Click options from the fields of a dataclass.\n\n\nbuild_command\nBuild command list from base command and options.\n\n\ndownload_file\nDownload a single file and return its processing status.\n\n\nfetch_from_github\nSync files from a specific directory in the GitHub repository.\n\n\nfilter_none_kwargs\nWraps function to remove None-valued kwargs.\n\n\nload_model_and_tokenizer\nHelper function for loading a model, tokenizer, and processor specified in the given axolotl\n\n\nstrip_optional_type\nExtracts the non-None type from an Optional / Union type.\n\n\n\n\n\ncli.utils.add_options_from_config(config_class)\nCreate Click options from the fields of a Pydantic model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig_class\nType[BaseModel]\nPyDantic model with fields to parse from the CLI\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nCallable\nFunction decorator for Axolotl CLI command.\n\n\n\n\n\n\n\ncli.utils.add_options_from_dataclass(config_class)\nCreate Click options from the fields of a dataclass.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig_class\nType[Any]\nDataclass with fields to parse from the CLI.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nCallable\nFunction decorator for Axolotl CLI command.\n\n\n\n\n\n\n\ncli.utils.build_command(base_cmd, options)\nBuild command list from base command and options.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbase_cmd\nlist[str]\nCommand without options.\nrequired\n\n\noptions\ndict[str, Any]\nOptions to parse and append to base command.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[str]\nList of strings giving shell command.\n\n\n\n\n\n\n\ncli.utils.download_file(file_info, raw_base_url, dest_path, dir_prefix)\nDownload a single file and return its processing status.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nfile_info\ntuple\nTuple of (file_path, remote_sha).\nrequired\n\n\nraw_base_url\nstr\nBase URL for raw GitHub content.\nrequired\n\n\ndest_path\nPath\nLocal destination directory.\nrequired\n\n\ndir_prefix\nstr\nDirectory prefix to filter files.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[str, str]\nTuple of (file_path, status) where status is ‘new’, ‘updated’, or ‘unchanged’.\n\n\n\n\n\n\n\ncli.utils.fetch_from_github(dir_prefix, dest_dir=None, max_workers=5)\nSync files from a specific directory in the GitHub repository.\nOnly downloads files that don’t exist locally or have changed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndir_prefix\nstr\nDirectory prefix to filter files (e.g., ‘examples/’, ‘deepspeed_configs/’).\nrequired\n\n\ndest_dir\nstr | None\nLocal destination directory.\nNone\n\n\nmax_workers\nint\nMaximum number of concurrent downloads.\n5\n\n\n\n\n\n\n\ncli.utils.filter_none_kwargs(func)\nWraps function to remove None-valued kwargs.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nfunc\nCallable\nFunction to wrap.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nCallable\nWrapped function.\n\n\n\n\n\n\n\ncli.utils.load_model_and_tokenizer(cfg, inference=False)\nHelper function for loading a model, tokenizer, and processor specified in the given axolotl\nconfig.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ninference\nbool\nBoolean denoting inference mode.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PreTrainedModel, PreTrainedTokenizer | PreTrainedTokenizerFast | Any, ProcessorMixin | None]\nTuple of (PreTrainedModel, PreTrainedTokenizer, ProcessorMixin).\n\n\n\n\n\n\n\ncli.utils.strip_optional_type(field_type)\nExtracts the non-None type from an Optional / Union type.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nfield_type\ntype | str | None\nType of field for Axolotl CLI command.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\n\nIf the input type is Union[T, None] or Optional[T], returns T. Otherwise returns the input type unchanged."
  },
  {
    "objectID": "docs/multi-gpu.html",
    "href": "docs/multi-gpu.html",
    "title": "Multi-GPU",
    "section": "",
    "text": "This guide covers advanced training configurations for multi-GPU setups using Axolotl.",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/multi-gpu.html#sec-overview",
    "href": "docs/multi-gpu.html#sec-overview",
    "title": "Multi-GPU",
    "section": "1 Overview",
    "text": "1 Overview\nAxolotl supports several methods for multi-GPU training:\n\nDeepSpeed (recommended)\nFSDP (Fully Sharded Data Parallel)\nSequence parallelism\nFSDP + QLoRA",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/multi-gpu.html#sec-deepspeed",
    "href": "docs/multi-gpu.html#sec-deepspeed",
    "title": "Multi-GPU",
    "section": "2 DeepSpeed",
    "text": "2 DeepSpeed\nDeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.\n\n2.1 Configuration\nAdd to your YAML config:\ndeepspeed: deepspeed_configs/zero1.json\n\n\n2.2 Usage\n# Fetch deepspeed configs (if not already present)\naxolotl fetch deepspeed_configs\n\n# Passing arg via config\naxolotl train config.yml\n\n# Passing arg via cli\naxolotl train config.yml --deepspeed deepspeed_configs/zero1.json\n\n\n2.3 ZeRO Stages\nWe provide default configurations for:\n\nZeRO Stage 1 (zero1.json)\nZeRO Stage 1 with torch compile (zero1_torch_compile.json)\nZeRO Stage 2 (zero2.json)\nZeRO Stage 3 (zero3.json)\nZeRO Stage 3 with bf16 (zero3_bf16.json)\nZeRO Stage 3 with bf16 and CPU offload params(zero3_bf16_cpuoffload_params.json)\nZeRO Stage 3 with bf16 and CPU offload params and optimizer (zero3_bf16_cpuoffload_all.json)\n\n\n\n\n\n\n\nTip\n\n\n\nChoose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.\nStart from Stage 1 -> Stage 2 -> Stage 3.",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/multi-gpu.html#sec-fsdp",
    "href": "docs/multi-gpu.html#sec-fsdp",
    "title": "Multi-GPU",
    "section": "3 FSDP",
    "text": "3 FSDP\n\n3.1 Basic FSDP Configuration\nfsdp:\n  - full_shard\n  - auto_wrap\nfsdp_config:\n  fsdp_offload_params: true\n  fsdp_state_dict_type: FULL_STATE_DICT\n  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/multi-gpu.html#sec-sequence-parallelism",
    "href": "docs/multi-gpu.html#sec-sequence-parallelism",
    "title": "Multi-GPU",
    "section": "4 Sequence parallelism",
    "text": "4 Sequence parallelism\nWe support sequence parallelism (SP) via the\nring-flash-attention project. This\nallows one to split up sequences across GPUs, which is useful in the event that a\nsingle sequence causes OOM errors during model training.\nSee our dedicated guide for more information.\n\n4.1 FSDP + QLoRA\nFor combining FSDP with QLoRA, see our dedicated guide.",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/multi-gpu.html#sec-performance",
    "href": "docs/multi-gpu.html#sec-performance",
    "title": "Multi-GPU",
    "section": "5 Performance Optimization",
    "text": "5 Performance Optimization\n\n5.1 Liger Kernel Integration\nPlease see docs for more info.",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/multi-gpu.html#sec-troubleshooting",
    "href": "docs/multi-gpu.html#sec-troubleshooting",
    "title": "Multi-GPU",
    "section": "6 Troubleshooting",
    "text": "6 Troubleshooting\n\n6.1 NCCL Issues\nFor NCCL-related problems, see our NCCL troubleshooting guide.\n\n\n6.2 Common Problems\n\nMemory IssuesTraining Instability\n\n\n\nReduce micro_batch_size\nReduce eval_batch_size\nAdjust gradient_accumulation_steps\nConsider using a higher ZeRO stage\n\n\n\n\nStart with DeepSpeed ZeRO-2\nMonitor loss values\nCheck learning rates\n\n\n\n\nFor more detailed troubleshooting, see our debugging guide.",
    "crumbs": [
      "Deployments",
      "Multi-GPU"
    ]
  },
  {
    "objectID": "docs/debugging.html",
    "href": "docs/debugging.html",
    "title": "Debugging",
    "section": "",
    "text": "This document provides some tips and tricks for debugging Axolotl. It also provides an example configuration for debugging with VSCode. A good debugging setup is essential to understanding how Axolotl code works behind the scenes.",
    "crumbs": [
      "Troubleshooting",
      "Debugging"
    ]
  },
  {
    "objectID": "docs/debugging.html#table-of-contents",
    "href": "docs/debugging.html#table-of-contents",
    "title": "Debugging",
    "section": "Table of Contents",
    "text": "Table of Contents\n\nGeneral Tips\nDebugging with VSCode\n\nBackground\nConfiguration\nCustomizing your debugger\nVideo Tutorial\n\nDebugging With Docker\n\nSetup\nAttach To Container\nVideo - Attaching To Docker On Remote Host",
    "crumbs": [
      "Troubleshooting",
      "Debugging"
    ]
  },
  {
    "objectID": "docs/debugging.html#general-tips",
    "href": "docs/debugging.html#general-tips",
    "title": "Debugging",
    "section": "General Tips",
    "text": "General Tips\nWhile debugging it’s helpful to simplify your test scenario as much as possible. Here are some tips for doing so:\n\n[!Important]\nAll of these tips are incorporated into the example configuration for debugging with VSCode below.\n\n\nMake sure you are using the latest version of axolotl: This project changes often and bugs get fixed fast. Check your git branch and make sure you have pulled the latest changes from main.\nEliminate concurrency: Restrict the number of processes to 1 for both training and data preprocessing:\n\nSet CUDA_VISIBLE_DEVICES to a single GPU, ex: export CUDA_VISIBLE_DEVICES=0.\nSet dataset_processes: 1 in your axolotl config or run the training command with --dataset_processes=1.\n\nUse a small dataset: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure sample_packing: False and eval_sample_packing: False to avoid errors. If you are in a pinch and don’t have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training. For example, to shard the dataset into 20 pieces, add the following to your axolotl config):\ndatasets:\n    ...\n    shards: 20\nUse a small model: A good example of a small model is TinyLlama/TinyLlama-1.1B-Chat-v1.0.\nMinimize iteration time: Make sure the training loop finishes as fast as possible, with these settings.\n\nmicro_batch_size: 1\nmax_steps: 1\nval_set_size: 0\n\nClear Caches: Axolotl caches certain steps and so does the underlying HuggingFace trainer. You may want to clear some of these caches when debugging.\n\nData preprocessing: When debugging data preprocessing, which includes prompt template formation, you may want to delete the directory set in dataset_prepared_path: in your axolotl config. If you didn’t set this value, the default is last_run_prepared.\nHF Hub: If you are debugging data preprocessing, you should clear the relevant HF cache HuggingFace cache, by deleting the appropriate ~/.cache/huggingface/datasets/... folder(s).\nThe recommended approach is to redirect all outputs and caches to a temporary folder and delete selected subfolders before each run. This is demonstrated in the example configuration below.",
    "crumbs": [
      "Troubleshooting",
      "Debugging"
    ]
  },
  {
    "objectID": "docs/debugging.html#debugging-with-vscode",
    "href": "docs/debugging.html#debugging-with-vscode",
    "title": "Debugging",
    "section": "Debugging with VSCode",
    "text": "Debugging with VSCode\n\nBackground\nThe below example shows how to configure VSCode to debug data preprocessing of the chat_template format. This is the format used when you have the following in your axolotl config:\ndatasets:\n  - path: <path to your chat_template formatted dataset> # example on HF Hub: fozziethebeat/alpaca_messages_2k_test\n    type: chat_template\n\n[!Important]\nIf you are already familiar with advanced VSCode debugging, you can skip the below explanation and look at the files .vscode/launch.json and .vscode/tasks.json for an example configuration.\n\n\n[!Tip]\nIf you prefer to watch a video, rather than read, you can skip to the video tutorial below (but doing both is recommended).\n\n\n\nSetup\nMake sure you have an editable install of Axolotl, which ensures that changes you make to the code are reflected at runtime. Run the following commands from the root of this project:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\nRemote Hosts\nIf you developing on a remote host, you can easily use VSCode to debug remotely. To do so, you will need to follow this remote - SSH guide. You can also see the video below on Docker and Remote SSH debugging.\n\n\n\nConfiguration\nThe easiest way to get started is to modify the .vscode/launch.json file in this project. This is just an example configuration, so you may need to modify or copy it to suit your needs.\nFor example, to mimic the command cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml, you would use the below configuration1. Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to devtools and set the env variable HF_HOME to a temporary folder that is later partially deleted. This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.\n// .vscode/launch.json\n{\n    \"version\": \"0.2.0\",\n    \"configurations\": [\n        {\n            \"name\": \"Debug axolotl prompt - chat_template\",\n            \"type\": \"python\",\n            \"module\": \"accelerate.commands.launch\",\n            \"request\": \"launch\",\n            \"args\": [\n                \"-m\", \"axolotl.cli.train\", \"dev_chat_template.yml\",\n                // The flags below simplify debugging by overriding the axolotl config\n                // with the debugging tips above.  Modify as needed.\n                \"--dataset_processes=1\",      // limits data preprocessing to one process\n                \"--max_steps=1\",              // limits training to just one step\n                \"--batch_size=1\",             // minimizes batch size\n                \"--micro_batch_size=1\",       // minimizes batch size\n                \"--val_set_size=0\",           // disables validation\n                \"--sample_packing=False\",     // disables sample packing which is necessary for small datasets\n                \"--eval_sample_packing=False\",// disables sample packing on eval set\n                \"--dataset_prepared_path=temp_debug/axolotl_outputs/data\", // send data outputs to a temp folder\n                \"--output_dir=temp_debug/axolotl_outputs/model\" // send model outputs to a temp folder\n                ],\n            \"console\": \"integratedTerminal\",      // show output in the integrated terminal\n            \"cwd\": \"${workspaceFolder}/devtools\", // set working directory to devtools from the root of the project\n            \"justMyCode\": true,                   // step through only axolotl code\n            \"env\": {\"CUDA_VISIBLE_DEVICES\": \"0\",  // Since we aren't doing distributed training, we need to limit to one GPU\n                    \"HF_HOME\": \"${workspaceFolder}/devtools/temp_debug/.hf-cache\"}, // send HF cache to a temp folder\n            \"preLaunchTask\": \"cleanup-for-dataprep\", // delete temp folders (see below)\n        }\n    ]\n}\nAdditional notes about this configuration:\n\nThe argument justMyCode is set to true such that you step through only the axolotl code. If you want to step into dependencies, set this to false.\nThe preLaunchTask: cleanup-for-dataprep is defined in .vscode/tasks.json and is used to delete the following folders before debugging, which is essential to ensure that the data pre-processing code is run from scratch:\n\n./devtools/temp_debug/axolotl_outputs\n./devtools/temp_debug/.hf-cache/datasets\n\n\n\n[!Tip]\nYou may not want to delete these folders. For example, if you are debugging model training instead of data pre-processing, you may NOT want to delete the cache or output folders. You may also need to add additional tasks to the tasks.json file depending on your use case.\n\nBelow is the ./vscode/tasks.json file that defines the cleanup-for-dataprep task. This task is run before each debugging session when you use the above configuration. Note how there are two tasks that delete the two folders mentioned above. The third task cleanup-for-dataprep is a composite task that combines the two tasks. A composite task is necessary because VSCode does not allow you to specify multiple tasks in the preLaunchTask argument of the launch.json file.\n// .vscode/tasks.json\n// this file is used by launch.json\n{\n    \"version\": \"2.0.0\",\n    \"tasks\": [\n      // this task changes into the devtools directory and deletes the temp_debug/axolotl_outputs folder\n      {\n        \"label\": \"delete-outputs\",\n        \"type\": \"shell\",\n        \"command\": \"rm -rf temp_debug/axolotl_outputs\",\n        \"options\":{ \"cwd\": \"${workspaceFolder}/devtools\"},\n        \"problemMatcher\": []\n      },\n      // this task changes into the devtools directory and deletes the `temp_debug/.hf-cache/datasets` folder\n      {\n        \"label\": \"delete-temp-hf-dataset-cache\",\n        \"type\": \"shell\",\n        \"command\": \"rm -rf temp_debug/.hf-cache/datasets\",\n        \"options\":{ \"cwd\": \"${workspaceFolder}/devtools\"},\n        \"problemMatcher\": []\n      },\n        // this task combines the two tasks above\n      {\n       \"label\": \"cleanup-for-dataprep\",\n       \"dependsOn\": [\"delete-outputs\", \"delete-temp-hf-dataset-cache\"],\n      }\n    ]\n}\n\n\nCustomizing your debugger\nYour debugging use case may differ from the example above. The easiest thing to do is to put your own axolotl config in the devtools folder and modify the launch.json file to use your config. You may also want to modify the preLaunchTask to delete different folders or not delete anything at all.\n\n\nVideo Tutorial\nThe following video tutorial walks through the above configuration and demonstrates how to debug with VSCode, (click the image below to watch):\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl w/VSCode",
    "crumbs": [
      "Troubleshooting",
      "Debugging"
    ]
  },
  {
    "objectID": "docs/debugging.html#debugging-with-docker",
    "href": "docs/debugging.html#debugging-with-docker",
    "title": "Debugging",
    "section": "Debugging With Docker",
    "text": "Debugging With Docker\nUsing official Axolotl Docker images is a great way to debug your code, and is a very popular way to use Axolotl. Attaching VSCode to Docker takes a few more steps.\n\nSetup\nOn the host that is running axolotl (ex: if you are using a remote host), clone the axolotl repo and change your current directory to the root:\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\n\n[!Tip]\nIf you already have axolotl cloned on your host, make sure you have the latest changes and change into the root of the project.\n\nNext, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:2\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-py3.10-cu118-2.0.1\n\n[!Tip]\nTo understand which containers are available, see the Docker section of the README and the DockerHub repo. For details of how the Docker containers are built, see axolotl’s Docker CI builds.\n\nYou will now be in the container. Next, perform an editable install of Axolotl:\npip3 install packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\nAttach To Container\nNext, if you are using a remote host, Remote into this host with VSCode. If you are using a local host, you can skip this step.\nNext, select Dev Containers: Attach to Running Container... using the command palette (CMD + SHIFT + P) in VSCode. You will be prompted to select a container to attach to. Select the container you just created. You will now be in the container with a working directory that is at the root of the project. Any changes you make to the code will be reflected both in the container and on the host.\nNow you are ready to debug as described above (see Debugging with VSCode).\n\n\nVideo - Attaching To Docker On Remote Host\nHere is a short video that demonstrates how to attach to a Docker container on a remote host:\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl Part 2: Attaching to Docker on a Remote Host",
    "crumbs": [
      "Troubleshooting",
      "Debugging"
    ]
  },
  {
    "objectID": "docs/debugging.html#footnotes",
    "href": "docs/debugging.html#footnotes",
    "title": "Debugging",
    "section": "Footnotes",
    "text": "Footnotes\n\n\nThe config actually mimics the command CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml, but this is the same thing.↩︎\nMany of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit. You can read more about these flags here.↩︎",
    "crumbs": [
      "Troubleshooting",
      "Debugging"
    ]
  },
  {
    "objectID": "TODO.html",
    "href": "TODO.html",
    "title": "todo list",
    "section": "",
    "text": "[] Validation of parameters for combinations that won’t work\n\n\n\n\nFSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203\nadamw_bnb_8bit doesn’t play well with FSDP offload"
  },
  {
    "objectID": "TODO.html#things-that-are-known-not-to-work",
    "href": "TODO.html#things-that-are-known-not-to-work",
    "title": "todo list",
    "section": "",
    "text": "FSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203\nadamw_bnb_8bit doesn’t play well with FSDP offload"
  },
  {
    "objectID": "docs/amd_hpc.html",
    "href": "docs/amd_hpc.html",
    "title": "AMD GPUs on HPC Systems",
    "section": "",
    "text": "This guide provides step-by-step instructions for installing and configuring Axolotl on a High-Performance Computing (HPC) environment equipped with AMD GPUs.",
    "crumbs": [
      "Deployments",
      "AMD GPUs on HPC Systems"
    ]
  },
  {
    "objectID": "docs/amd_hpc.html#setup",
    "href": "docs/amd_hpc.html#setup",
    "title": "AMD GPUs on HPC Systems",
    "section": "Setup",
    "text": "Setup\n\n1. Install Python\nWe recommend using Miniforge, a minimal conda-based Python distribution:\ncurl -L -O \"https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh\"\nbash Miniforge3-$(uname)-$(uname -m).sh\n\n\n2. Configure Python Environment\nAdd Python to your PATH and ensure it’s available at login:\necho 'export PATH=~/miniforge3/bin:$PATH' >> ~/.bashrc\necho 'if [ -f ~/.bashrc ]; then . ~/.bashrc; fi' >> ~/.bash_profile\n\n\n3. Load AMD GPU Software\nLoad the ROCm module:\nmodule load rocm/5.7.1\nNote: The specific module name and version may vary depending on your HPC system. Consult your system documentation for the correct module name.\n\n\n4. Install PyTorch\nInstall PyTorch with ROCm support:\npip install -U torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.7 --force-reinstall\n\n\n5. Install Flash Attention\nClone and install the Flash Attention repository:\ngit clone --recursive https://github.com/ROCmSoftwarePlatform/flash-attention.git\nexport GPU_ARCHS=\"gfx90a\"\ncd flash-attention\nexport PYTHON_SITE_PACKAGES=$(python -c 'import site; print(site.getsitepackages()[0])')\npatch \"${PYTHON_SITE_PACKAGES}/torch/utils/hipify/hipify_python.py\" hipify_patch.patch\npip install --no-build-isolation .\n\n\n6. Install Axolotl\nClone and install Axolotl:\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\npip install packaging ninja\npip install --no-build-isolation -e .\n\n\n7. Apply xformers Workaround\nxformers appears to be incompatible with ROCm. Apply the following workarounds:\n- Edit $HOME/packages/axolotl/src/axolotl/monkeypatch/llama_attn_hijack_flash.py modifying the code to always return False for SwiGLU availability from xformers.\n- Edit $HOME/miniforge3/lib/python3.10/site-packages/xformers/ops/swiglu_op.py replacing the “SwiGLU” function with a pass statement.\n\n\n8. Prepare Job Submission Script\nCreate a script for job submission using your HPC’s particular software (e.g. Slurm, PBS). Include necessary environment setup and the command to run Axolotl training. If the compute node(s) do(es) not have internet access, it is recommended to include\nexport TRANSFORMERS_OFFLINE=1\nexport HF_DATASETS_OFFLINE=1\n\n\n9. Download Base Model\nDownload a base model using the Hugging Face CLI:\nhuggingface-cli download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B\n\n\n10. Create Axolotl Configuration\nCreate an Axolotl configuration file (YAML format) tailored to your specific training requirements and dataset. Use FSDP for multi-node training.\nNote: Deepspeed did not work at the time of testing. However, if anyone managed to get it working, please let us know.\n\n\n11. Preprocess Data\nRun preprocessing on the login node:\nCUDA_VISIBLE_DEVICES=\"\" python -m axolotl.cli.preprocess /path/to/your/config.yaml\n\n\n12. Train\nYou are now ready to submit your previously prepared job script. 🚂",
    "crumbs": [
      "Deployments",
      "AMD GPUs on HPC Systems"
    ]
  },
  {
    "objectID": "docs/dataset_loading.html",
    "href": "docs/dataset_loading.html",
    "title": "Dataset Loading",
    "section": "",
    "text": "Datasets can be loaded in a number of different ways depending on the how it is saved (the extension of the file) and where it is stored.",
    "crumbs": [
      "How To Guides",
      "Dataset Loading"
    ]
  },
  {
    "objectID": "docs/dataset_loading.html#overview",
    "href": "docs/dataset_loading.html#overview",
    "title": "Dataset Loading",
    "section": "",
    "text": "Datasets can be loaded in a number of different ways depending on the how it is saved (the extension of the file) and where it is stored.",
    "crumbs": [
      "How To Guides",
      "Dataset Loading"
    ]
  },
  {
    "objectID": "docs/dataset_loading.html#loading-datasets",
    "href": "docs/dataset_loading.html#loading-datasets",
    "title": "Dataset Loading",
    "section": "Loading Datasets",
    "text": "Loading Datasets\nWe use the datasets library to load datasets and a mix of load_dataset and load_from_disk to load them.\nYou may recognize the similar named configs between load_dataset and the datasets section of the config file.\ndatasets:\n  - path:\n    name:\n    data_files:\n    split:\n    revision:\n    trust_remote_code:\n\n\n\n\n\n\nTip\n\n\n\nDo not feel overwhelmed by the number of options here. A lot of them are optional. In fact, the most common config to use would be path and sometimes data_files.\n\n\nThis matches the API of datasets.load_dataset, so if you’re familiar with that, you will feel right at home.\nFor HuggingFace’s guide to load different dataset types, see here.\nFor full details on the config, see config.qmd.\n\n\n\n\n\n\nNote\n\n\n\nYou can set multiple datasets in the config file by more than one entry under datasets.\ndatasets:\n  - path: /path/to/your/dataset\n  - path: /path/to/your/other/dataset\n\n\n\nLocal dataset\n\nFiles\nTo load a JSON file, you would do something like this:\nfrom datasets import load_dataset\n\ndataset = load_dataset(\"json\", data_files=\"data.json\")\nWhich translates to the following config:\ndatasets:\n  - path: data.json\n    ds_type: json\nIn the example above, it can be seen that we can just point the path to the file or directory along with the ds_type to load the dataset.\nThis works for CSV, JSON, Parquet, and Arrow files.\n\n\n\n\n\n\nTip\n\n\n\nIf path points to a file and ds_type is not specified, we will automatically infer the dataset type from the file extension, so you could omit ds_type if you’d like.\n\n\n\n\nDirectory\nIf you’re loading a directory, you can point the path to the directory.\nThen, you have two options:\n\nLoading entire directory\nYou do not need any additional configs.\nWe will attempt to load in the following order:\n- datasets saved with datasets.save_to_disk\n- loading entire directory of files (such as with parquet/arrow files)\ndatasets:\n  - path: /path/to/your/directory\n\n\nLoading specific files in directory\nProvide data_files with a list of files to load.\ndatasets:\n    # single file\n  - path: /path/to/your/directory\n    ds_type: csv\n    data_files: file1.csv\n\n    # multiple files\n  - path: /path/to/your/directory\n    ds_type: json\n    data_files:\n      - file1.jsonl\n      - file2.jsonl\n\n    # multiple files for parquet\n  - path: /path/to/your/directory\n    ds_type: parquet\n    data_files:\n      - file1.parquet\n      - file2.parquet\n\n\n\n\nHuggingFace Hub\nThe method you use to load the dataset depends on how the dataset was created, whether a folder was uploaded directly or a HuggingFace Dataset was pushed.\n\n\n\n\n\n\nNote\n\n\n\nIf you’re using a private dataset, you will need to enable the hf_use_auth_token flag in the root-level of the config file.\n\n\n\nFolder uploaded\nThis would mean that the dataset is a single file or file(s) uploaded to the Hub.\ndatasets:\n  - path: org/dataset-name\n    data_files:\n      - file1.jsonl\n      - file2.jsonl\n\n\nHuggingFace Dataset\nThis means that the dataset is created as a HuggingFace Dataset and pushed to the Hub via datasets.push_to_hub.\ndatasets:\n  - path: org/dataset-name\n\n\n\n\n\n\nNote\n\n\n\nThere are some other configs which may be required like name, split, revision, trust_remote_code, etc depending on the dataset.\n\n\n\n\n\nRemote Filesystems\nVia the storage_options config under load_dataset, you can load datasets from remote filesystems like S3, GCS, Azure, and OCI.\n\n\n\n\n\n\nWarning\n\n\n\nThis is currently experimental. Please let us know if you run into any issues!\n\n\nThe only difference between the providers is that you need to prepend the path with the respective protocols.\ndatasets:\n    # Single file\n  - path: s3://bucket-name/path/to/your/file.jsonl\n\n    # Directory\n  - path: s3://bucket-name/path/to/your/directory\nFor directory, we load via load_from_disk.\n\nS3\nPrepend the path with s3://.\nThe credentials are pulled in the following order:\n\nAWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and AWS_SESSION_TOKEN environment variables\nfrom the ~/.aws/credentials file\nfor nodes on EC2, the IAM metadata provider\n\n\n\n\n\n\n\nNote\n\n\n\nWe assume you have credentials setup and not using anonymous access. If you want to use anonymous access, let us know! We may have to open a config option for this.\n\n\nOther environment variables that can be set can be found in boto3 docs\n\n\nGCS\nPrepend the path with gs:// or gcs://.\nThe credentials are loaded in the following order:\n\ngcloud credentials\nfor nodes on GCP, the google metadata service\nanonymous access\n\n\n\nAzure\n\nGen 1\nPrepend the path with adl://.\nEnsure you have the following environment variables set:\n\nAZURE_STORAGE_TENANT_ID\nAZURE_STORAGE_CLIENT_ID\nAZURE_STORAGE_CLIENT_SECRET\n\n\n\nGen 2\nPrepend the path with abfs:// or az://.\nEnsure you have the following environment variables set:\n\nAZURE_STORAGE_ACCOUNT_NAME\nAZURE_STORAGE_ACCOUNT_KEY\n\nOther environment variables that can be set can be found in adlfs docs\n\n\n\nOCI\nPrepend the path with oci://.\nIt would attempt to read in the following order:\n\nOCIFS_IAM_TYPE, OCIFS_CONFIG_LOCATION, and OCIFS_CONFIG_PROFILE environment variables\nwhen on OCI resource, resource principal\n\nOther environment variables:\n\nOCI_REGION_METADATA\n\nPlease see the ocifs docs.\n\n\n\nHTTPS\nThe path should start with https://.\ndatasets:\n  - path: https://path/to/your/dataset/file.jsonl\nThis must be publically accessible.",
    "crumbs": [
      "How To Guides",
      "Dataset Loading"
    ]
  },
  {
    "objectID": "docs/dataset_loading.html#next-steps",
    "href": "docs/dataset_loading.html#next-steps",
    "title": "Dataset Loading",
    "section": "Next steps",
    "text": "Next steps\nNow that you know how to load datasets, you can learn more on how to load your specific dataset format into your target output format dataset formats docs.",
    "crumbs": [
      "How To Guides",
      "Dataset Loading"
    ]
  },
  {
    "objectID": "docs/api/cli.merge_sharded_fsdp_weights.html",
    "href": "docs/api/cli.merge_sharded_fsdp_weights.html",
    "title": "cli.merge_sharded_fsdp_weights",
    "section": "",
    "text": "cli.merge_sharded_fsdp_weights\nCLI to merge sharded FSDP model checkpoints into a single combined checkpoint.\n\n\n\n\n\nName\nDescription\n\n\n\n\nBFloat16CastPlanner\nA custom planner to cast tensors to bfloat16 on the fly during loading.\n\n\n\n\n\ncli.merge_sharded_fsdp_weights.BFloat16CastPlanner()\nA custom planner to cast tensors to bfloat16 on the fly during loading.\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls merge_fsdp_weights.\n\n\nmerge_fsdp_weights\nMerge the weights from sharded FSDP model checkpoints into a single combined checkpoint. Should be used if\n\n\n\n\n\ncli.merge_sharded_fsdp_weights.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls merge_fsdp_weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.merge_sharded_fsdp_weights.merge_fsdp_weights(\n    checkpoint_dir,\n    output_path,\n    safe_serialization=False,\n    remove_checkpoint_dir=False,\n)\nMerge the weights from sharded FSDP model checkpoints into a single combined checkpoint. Should be used if\nSHARDED_STATE_DICT was used for the model. Weights will be saved to {output_path}/model.safetensors if\nsafe_serialization else pytorch_model.bin.\nNote: this is a CPU-bound process.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncheckpoint_dir\nstr\nThe directory containing the FSDP checkpoints (can be either the model or optimizer).\nrequired\n\n\noutput_path\nstr\nThe path to save the merged checkpoint.\nrequired\n\n\nsafe_serialization\nbool, optional, defaults to True\nWhether to save the merged weights with safetensors (recommended).\nFalse\n\n\nremove_checkpoint_dir\nbool, optional, defaults to False\nWhether to remove the checkpoint directory after merging.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf torch version < 2.3.0, or if checkpoint_dir does not exist."
  },
  {
    "objectID": "docs/api/cli.merge_sharded_fsdp_weights.html#classes",
    "href": "docs/api/cli.merge_sharded_fsdp_weights.html#classes",
    "title": "cli.merge_sharded_fsdp_weights",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nBFloat16CastPlanner\nA custom planner to cast tensors to bfloat16 on the fly during loading.\n\n\n\n\n\ncli.merge_sharded_fsdp_weights.BFloat16CastPlanner()\nA custom planner to cast tensors to bfloat16 on the fly during loading."
  },
  {
    "objectID": "docs/api/cli.merge_sharded_fsdp_weights.html#functions",
    "href": "docs/api/cli.merge_sharded_fsdp_weights.html#functions",
    "title": "cli.merge_sharded_fsdp_weights",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls merge_fsdp_weights.\n\n\nmerge_fsdp_weights\nMerge the weights from sharded FSDP model checkpoints into a single combined checkpoint. Should be used if\n\n\n\n\n\ncli.merge_sharded_fsdp_weights.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls merge_fsdp_weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.merge_sharded_fsdp_weights.merge_fsdp_weights(\n    checkpoint_dir,\n    output_path,\n    safe_serialization=False,\n    remove_checkpoint_dir=False,\n)\nMerge the weights from sharded FSDP model checkpoints into a single combined checkpoint. Should be used if\nSHARDED_STATE_DICT was used for the model. Weights will be saved to {output_path}/model.safetensors if\nsafe_serialization else pytorch_model.bin.\nNote: this is a CPU-bound process.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncheckpoint_dir\nstr\nThe directory containing the FSDP checkpoints (can be either the model or optimizer).\nrequired\n\n\noutput_path\nstr\nThe path to save the merged checkpoint.\nrequired\n\n\nsafe_serialization\nbool, optional, defaults to True\nWhether to save the merged weights with safetensors (recommended).\nFalse\n\n\nremove_checkpoint_dir\nbool, optional, defaults to False\nWhether to remove the checkpoint directory after merging.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf torch version < 2.3.0, or if checkpoint_dir does not exist."
  },
  {
    "objectID": "docs/api/monkeypatch.attention.mllama.html",
    "href": "docs/api/monkeypatch.attention.mllama.html",
    "title": "monkeypatch.attention.mllama",
    "section": "",
    "text": "monkeypatch.attention.mllama\nMonkeypatch for Vision Llama for FA2 support\n\n\n\n\n\nName\nDescription\n\n\n\n\nMllamaTextCrossFlashAttention2\nMllama flash cross-attention module. This module inherits from MllamaTextCrossAttention and\n\n\nMllamaTextSelfFlashAttention2\nMllama flash self-attention module. This module inherits from MllamaTextSelfAttention and\n\n\n\n\n\nmonkeypatch.attention.mllama.MllamaTextCrossFlashAttention2(*args, **kwargs)\nMllama flash cross-attention module. This module inherits from MllamaTextCrossAttention and\nimplements the forward pass using Flash Attention for improved performance.\n\n\n\nmonkeypatch.attention.mllama.MllamaTextSelfFlashAttention2(\n    config,\n    layer_idx,\n    *args,\n    **kwargs,\n)\nMllama flash self-attention module. This module inherits from MllamaTextSelfAttention and\nimplements the forward pass using Flash Attention for improved performance."
  },
  {
    "objectID": "docs/api/monkeypatch.attention.mllama.html#classes",
    "href": "docs/api/monkeypatch.attention.mllama.html#classes",
    "title": "monkeypatch.attention.mllama",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMllamaTextCrossFlashAttention2\nMllama flash cross-attention module. This module inherits from MllamaTextCrossAttention and\n\n\nMllamaTextSelfFlashAttention2\nMllama flash self-attention module. This module inherits from MllamaTextSelfAttention and\n\n\n\n\n\nmonkeypatch.attention.mllama.MllamaTextCrossFlashAttention2(*args, **kwargs)\nMllama flash cross-attention module. This module inherits from MllamaTextCrossAttention and\nimplements the forward pass using Flash Attention for improved performance.\n\n\n\nmonkeypatch.attention.mllama.MllamaTextSelfFlashAttention2(\n    config,\n    layer_idx,\n    *args,\n    **kwargs,\n)\nMllama flash self-attention module. This module inherits from MllamaTextSelfAttention and\nimplements the forward pass using Flash Attention for improved performance."
  },
  {
    "objectID": "docs/api/common.architectures.html",
    "href": "docs/api/common.architectures.html",
    "title": "common.architectures",
    "section": "",
    "text": "common.architectures\ncommon.architectures\nCommon architecture specific constants"
  },
  {
    "objectID": "docs/api/utils.schedulers.html",
    "href": "docs/api/utils.schedulers.html",
    "title": "utils.schedulers",
    "section": "",
    "text": "utils.schedulers\nModule for custom LRScheduler class\n\n\n\n\n\nName\nDescription\n\n\n\n\nInterpolatingLogScheduler\nA scheduler that interpolates learning rates in a logarithmic fashion\n\n\nRexLR\nReflected Exponential (REX) learning rate scheduler.\n\n\n\n\n\nutils.schedulers.InterpolatingLogScheduler(\n    optimizer,\n    num_steps,\n    min_lr,\n    max_lr,\n    last_epoch=-1,\n)\nA scheduler that interpolates learning rates in a logarithmic fashion\n\n\n\nutils.schedulers.RexLR(\n    optimizer,\n    max_lr,\n    min_lr,\n    total_steps=0,\n    num_warmup_steps=0,\n    last_step=0,\n)\nReflected Exponential (REX) learning rate scheduler.\n\nOriginal implementation: https://github.com/IvanVassi/REX_LR\nOriginal license: Apache 2.0\nBased on: https://arxiv.org/abs/2107.04197\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\noptimizer\ntorch.optim.Optimizer\nThe optimizer to schedule the learning rate for.\nrequired\n\n\nmax_lr\nfloat\nThe maximum learning rate.\nrequired\n\n\nmin_lr\nfloat\nThe minimum learning rate.\nrequired\n\n\ntotal_steps\nint\nThe total number of training steps.\n0\n\n\nnum_warmup_steps\nint\nThe number of warmup steps.\n0\n\n\nlast_step\nint\nThe index of last step.\n0\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_cosine_schedule_with_min_lr\n\n\n\nget_cosine_schedule_with_quadratic_warmup\nCreate a schedule with a learning rate that decreases following the values of the cosine function between the\n\n\nget_cosine_schedule_with_warmup_decay_constant\nImplementation of Continual Pre-Training of Large Language Models: How to (re)warm your model? (https://arxiv.org/pdf/2308.04014.pdf)\n\n\n\n\n\nutils.schedulers.get_cosine_schedule_with_min_lr(\n    optimizer,\n    num_warmup_steps,\n    num_training_steps,\n    min_lr_ratio=0.0,\n)\n\n\n\nlinear warmup from 0 -> max_lr over num_warmup_steps\ncosine learning rate annealing from max_lr -> min_lr over num_training_steps\n\n\n\n\n\nutils.schedulers.get_cosine_schedule_with_quadratic_warmup(\n    optimizer,\n    num_warmup_steps,\n    num_training_steps,\n    num_cycles=0.5,\n    last_epoch=-1,\n)\nCreate a schedule with a learning rate that decreases following the values of the cosine function between the\ninitial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the\ninitial lr set in the optimizer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\noptimizer\n[~torch.optim.Optimizer]\nThe optimizer for which to schedule the learning rate.\nrequired\n\n\nnum_warmup_steps\nint\nThe number of steps for the warmup phase.\nrequired\n\n\nnum_training_steps\nint\nThe total number of training steps.\nrequired\n\n\nnum_cycles\nfloat, optional, defaults to 0.5\nThe number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0 following a half-cosine).\n0.5\n\n\nlast_epoch\nint, optional, defaults to -1\nThe index of the last epoch when resuming training.\n-1\n\n\n\n\n\n\ntorch.optim.lr_scheduler.LambdaLR with the appropriate schedule.\n\n\n\n\nutils.schedulers.get_cosine_schedule_with_warmup_decay_constant(\n    optimizer,\n    num_warmup_steps,\n    num_training_steps,\n    constant_lr_ratio,\n    min_lr_ratio,\n    num_cycles=0.5,\n    last_epoch=-1,\n)\nImplementation of Continual Pre-Training of Large Language Models: How to (re)warm your model? (https://arxiv.org/pdf/2308.04014.pdf)\nCreate a schedule with a learning rate that decreases following the values of the cosine function between the\ninitial lr set in the optimizer to min_lr_ratio until num_training_steps * constant_lr_ratio, after constant_rate returns constant value of min_rate\n, after a warmup period during which it increases linearly between 0 and the initial lr set in the optimizer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\noptimizer\n[~torch.optim.Optimizer]\nThe optimizer for which to schedule the learning rate.\nrequired\n\n\nnum_warmup_steps\nint\nThe number of steps for the warmup phase.\nrequired\n\n\nnum_training_steps\nint\nThe total number of training steps.\nrequired\n\n\nconstant_lr_ratio\nfloat\n(float): The ratio of num_training_steps to decrease by cosine function.\nrequired\n\n\nmin_lr_ratio\nfloat\n(float): The ratio of maximum learning rate for cosine function to decay to minimum learning rate.                            | _required_ | | num_cycles         |float, *optional*, defaults to 0.5 | The number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0 following a half-cosine). |0.5| | last_epoch         |int, *optional*, defaults to -1    | The index of the last epoch when resuming training.                                                                            |-1`\n\n\n\n\n\n\n\ntorch.optim.lr_scheduler.LambdaLR with the appropriate schedule."
  },
  {
    "objectID": "docs/api/utils.schedulers.html#classes",
    "href": "docs/api/utils.schedulers.html#classes",
    "title": "utils.schedulers",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nInterpolatingLogScheduler\nA scheduler that interpolates learning rates in a logarithmic fashion\n\n\nRexLR\nReflected Exponential (REX) learning rate scheduler.\n\n\n\n\n\nutils.schedulers.InterpolatingLogScheduler(\n    optimizer,\n    num_steps,\n    min_lr,\n    max_lr,\n    last_epoch=-1,\n)\nA scheduler that interpolates learning rates in a logarithmic fashion\n\n\n\nutils.schedulers.RexLR(\n    optimizer,\n    max_lr,\n    min_lr,\n    total_steps=0,\n    num_warmup_steps=0,\n    last_step=0,\n)\nReflected Exponential (REX) learning rate scheduler.\n\nOriginal implementation: https://github.com/IvanVassi/REX_LR\nOriginal license: Apache 2.0\nBased on: https://arxiv.org/abs/2107.04197\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\noptimizer\ntorch.optim.Optimizer\nThe optimizer to schedule the learning rate for.\nrequired\n\n\nmax_lr\nfloat\nThe maximum learning rate.\nrequired\n\n\nmin_lr\nfloat\nThe minimum learning rate.\nrequired\n\n\ntotal_steps\nint\nThe total number of training steps.\n0\n\n\nnum_warmup_steps\nint\nThe number of warmup steps.\n0\n\n\nlast_step\nint\nThe index of last step.\n0"
  },
  {
    "objectID": "docs/api/utils.schedulers.html#functions",
    "href": "docs/api/utils.schedulers.html#functions",
    "title": "utils.schedulers",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nget_cosine_schedule_with_min_lr\n\n\n\nget_cosine_schedule_with_quadratic_warmup\nCreate a schedule with a learning rate that decreases following the values of the cosine function between the\n\n\nget_cosine_schedule_with_warmup_decay_constant\nImplementation of Continual Pre-Training of Large Language Models: How to (re)warm your model? (https://arxiv.org/pdf/2308.04014.pdf)\n\n\n\n\n\nutils.schedulers.get_cosine_schedule_with_min_lr(\n    optimizer,\n    num_warmup_steps,\n    num_training_steps,\n    min_lr_ratio=0.0,\n)\n\n\n\nlinear warmup from 0 -> max_lr over num_warmup_steps\ncosine learning rate annealing from max_lr -> min_lr over num_training_steps\n\n\n\n\n\nutils.schedulers.get_cosine_schedule_with_quadratic_warmup(\n    optimizer,\n    num_warmup_steps,\n    num_training_steps,\n    num_cycles=0.5,\n    last_epoch=-1,\n)\nCreate a schedule with a learning rate that decreases following the values of the cosine function between the\ninitial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the\ninitial lr set in the optimizer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\noptimizer\n[~torch.optim.Optimizer]\nThe optimizer for which to schedule the learning rate.\nrequired\n\n\nnum_warmup_steps\nint\nThe number of steps for the warmup phase.\nrequired\n\n\nnum_training_steps\nint\nThe total number of training steps.\nrequired\n\n\nnum_cycles\nfloat, optional, defaults to 0.5\nThe number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0 following a half-cosine).\n0.5\n\n\nlast_epoch\nint, optional, defaults to -1\nThe index of the last epoch when resuming training.\n-1\n\n\n\n\n\n\ntorch.optim.lr_scheduler.LambdaLR with the appropriate schedule.\n\n\n\n\nutils.schedulers.get_cosine_schedule_with_warmup_decay_constant(\n    optimizer,\n    num_warmup_steps,\n    num_training_steps,\n    constant_lr_ratio,\n    min_lr_ratio,\n    num_cycles=0.5,\n    last_epoch=-1,\n)\nImplementation of Continual Pre-Training of Large Language Models: How to (re)warm your model? (https://arxiv.org/pdf/2308.04014.pdf)\nCreate a schedule with a learning rate that decreases following the values of the cosine function between the\ninitial lr set in the optimizer to min_lr_ratio until num_training_steps * constant_lr_ratio, after constant_rate returns constant value of min_rate\n, after a warmup period during which it increases linearly between 0 and the initial lr set in the optimizer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\noptimizer\n[~torch.optim.Optimizer]\nThe optimizer for which to schedule the learning rate.\nrequired\n\n\nnum_warmup_steps\nint\nThe number of steps for the warmup phase.\nrequired\n\n\nnum_training_steps\nint\nThe total number of training steps.\nrequired\n\n\nconstant_lr_ratio\nfloat\n(float): The ratio of num_training_steps to decrease by cosine function.\nrequired\n\n\nmin_lr_ratio\nfloat\n(float): The ratio of maximum learning rate for cosine function to decay to minimum learning rate.                            | _required_ | | num_cycles         |float, *optional*, defaults to 0.5 | The number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0 following a half-cosine). |0.5| | last_epoch         |int, *optional*, defaults to -1    | The index of the last epoch when resuming training.                                                                            |-1`\n\n\n\n\n\n\n\ntorch.optim.lr_scheduler.LambdaLR with the appropriate schedule."
  },
  {
    "objectID": "docs/api/utils.collators.core.html",
    "href": "docs/api/utils.collators.core.html",
    "title": "utils.collators.core",
    "section": "",
    "text": "utils.collators.core\nutils.collators.core\nbasic shared collator constants"
  },
  {
    "objectID": "docs/api/utils.schemas.multimodal.html",
    "href": "docs/api/utils.schemas.multimodal.html",
    "title": "utils.schemas.multimodal",
    "section": "",
    "text": "utils.schemas.multimodal\nPydantic models for multimodal-related configuration\n\n\n\n\n\nName\nDescription\n\n\n\n\nMultiModalConfig\nMulti-modal configuration subset\n\n\n\n\n\nutils.schemas.multimodal.MultiModalConfig()\nMulti-modal configuration subset\n\n\n\n\n\nName\nDescription\n\n\n\n\nconvert_image_resize_algorithm\nConvert the image resize algorithm to a PIL.Image.Resampling enum.\n\n\n\n\n\nutils.schemas.multimodal.MultiModalConfig.convert_image_resize_algorithm(\n    image_resize_algorithm,\n)\nConvert the image resize algorithm to a PIL.Image.Resampling enum."
  },
  {
    "objectID": "docs/api/utils.schemas.multimodal.html#classes",
    "href": "docs/api/utils.schemas.multimodal.html#classes",
    "title": "utils.schemas.multimodal",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMultiModalConfig\nMulti-modal configuration subset\n\n\n\n\n\nutils.schemas.multimodal.MultiModalConfig()\nMulti-modal configuration subset\n\n\n\n\n\nName\nDescription\n\n\n\n\nconvert_image_resize_algorithm\nConvert the image resize algorithm to a PIL.Image.Resampling enum.\n\n\n\n\n\nutils.schemas.multimodal.MultiModalConfig.convert_image_resize_algorithm(\n    image_resize_algorithm,\n)\nConvert the image resize algorithm to a PIL.Image.Resampling enum."
  },
  {
    "objectID": "docs/api/utils.distributed.html",
    "href": "docs/api/utils.distributed.html",
    "title": "utils.distributed",
    "section": "",
    "text": "utils.distributed\nutility helpers for distributed checks\n\n\n\n\n\nName\nDescription\n\n\n\n\nbarrier\nActs as a barrier to wait for all processes. This ensures that all processes\n\n\ncleanup_distributed\nDestroy process group if torch distributed is initialized. Called in training early\n\n\ncompute_and_broadcast\nCompute a value using the function ‘fn’ only on the specified rank (default is 0).\n\n\ngather_from_all_ranks\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\n\n\ngather_scalar_from_all_ranks\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\n\n\nis_distributed\nCheck if distributed training is initialized.\n\n\nis_main_process\nCheck if the current process is the main process. If not in distributed mode,\n\n\nreduce_and_broadcast\nRun a callable ‘fn1’ on all ranks, gather the results, reduce them using ‘fn2’,\n\n\nzero_first\nruns the wrapped context so that rank 0 runs first before other ranks\n\n\n\n\n\nutils.distributed.barrier()\nActs as a barrier to wait for all processes. This ensures that all processes\nreach the barrier before proceeding further.\n\n\n\nutils.distributed.cleanup_distributed()\nDestroy process group if torch distributed is initialized. Called in training early\ntermination or when training successfully completes.\n\n\n\nutils.distributed.compute_and_broadcast(fn)\nCompute a value using the function ‘fn’ only on the specified rank (default is 0).\nThe value is then broadcasted to all other ranks.\nArgs:\n- fn (callable): A function that computes the value. This should not have any side effects.\n- rank (int, optional): The rank that computes the value. Default is 0.\nReturns:\n- The computed value (int or float).\n\n\n\nutils.distributed.gather_from_all_ranks(fn, world_size=1)\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\nArgs:\n- fn (callable): A function that computes the value. This should not have any side effects.\n- rank (int, optional): The rank that gathers the values. Default is 0.\n- world_size (int, optional): Total number of processes in the current distributed setup.\nReturns:\n- A list of computed values from all ranks if on the gathering rank, otherwise None.\n\n\n\nutils.distributed.gather_scalar_from_all_ranks(fn, world_size=1)\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\nArgs:\n- fn (callable): A function that computes the value. This should not have any side effects.\n- rank (int, optional): The rank that gathers the values. Default is 0.\n- world_size (int, optional): Total number of processes in the current distributed setup.\nReturns:\n- A list of computed values from all ranks if on the gathering rank, otherwise None.\n\n\n\nutils.distributed.is_distributed()\nCheck if distributed training is initialized.\n\n\n\nutils.distributed.is_main_process(use_environ=False)\nCheck if the current process is the main process. If not in distributed mode,\nalways return True.\nArgs:\n- use_environ (bool, optional): Use environment variable to determine main process.\nReturns:\n- bool: True if the current process is the main process, False otherwise.\n\n\n\nutils.distributed.reduce_and_broadcast(fn1, fn2)\nRun a callable ‘fn1’ on all ranks, gather the results, reduce them using ‘fn2’,\nand then broadcast the reduced result to all ranks.\nArgs:\n- fn1 (callable): A function that computes the value on each rank.\n- fn2 (callable): A reduction function that takes a list of values and returns a single value.\n- world_size (int, optional): Total number of processes in the current distributed setup.\nReturns:\n- The reduced and broadcasted value.\n\n\n\nutils.distributed.zero_first(is_main)\nruns the wrapped context so that rank 0 runs first before other ranks"
  },
  {
    "objectID": "docs/api/utils.distributed.html#functions",
    "href": "docs/api/utils.distributed.html#functions",
    "title": "utils.distributed",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nbarrier\nActs as a barrier to wait for all processes. This ensures that all processes\n\n\ncleanup_distributed\nDestroy process group if torch distributed is initialized. Called in training early\n\n\ncompute_and_broadcast\nCompute a value using the function ‘fn’ only on the specified rank (default is 0).\n\n\ngather_from_all_ranks\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\n\n\ngather_scalar_from_all_ranks\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\n\n\nis_distributed\nCheck if distributed training is initialized.\n\n\nis_main_process\nCheck if the current process is the main process. If not in distributed mode,\n\n\nreduce_and_broadcast\nRun a callable ‘fn1’ on all ranks, gather the results, reduce them using ‘fn2’,\n\n\nzero_first\nruns the wrapped context so that rank 0 runs first before other ranks\n\n\n\n\n\nutils.distributed.barrier()\nActs as a barrier to wait for all processes. This ensures that all processes\nreach the barrier before proceeding further.\n\n\n\nutils.distributed.cleanup_distributed()\nDestroy process group if torch distributed is initialized. Called in training early\ntermination or when training successfully completes.\n\n\n\nutils.distributed.compute_and_broadcast(fn)\nCompute a value using the function ‘fn’ only on the specified rank (default is 0).\nThe value is then broadcasted to all other ranks.\nArgs:\n- fn (callable): A function that computes the value. This should not have any side effects.\n- rank (int, optional): The rank that computes the value. Default is 0.\nReturns:\n- The computed value (int or float).\n\n\n\nutils.distributed.gather_from_all_ranks(fn, world_size=1)\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\nArgs:\n- fn (callable): A function that computes the value. This should not have any side effects.\n- rank (int, optional): The rank that gathers the values. Default is 0.\n- world_size (int, optional): Total number of processes in the current distributed setup.\nReturns:\n- A list of computed values from all ranks if on the gathering rank, otherwise None.\n\n\n\nutils.distributed.gather_scalar_from_all_ranks(fn, world_size=1)\nRun a callable ‘fn’ on all ranks and gather the results on the specified rank.\nArgs:\n- fn (callable): A function that computes the value. This should not have any side effects.\n- rank (int, optional): The rank that gathers the values. Default is 0.\n- world_size (int, optional): Total number of processes in the current distributed setup.\nReturns:\n- A list of computed values from all ranks if on the gathering rank, otherwise None.\n\n\n\nutils.distributed.is_distributed()\nCheck if distributed training is initialized.\n\n\n\nutils.distributed.is_main_process(use_environ=False)\nCheck if the current process is the main process. If not in distributed mode,\nalways return True.\nArgs:\n- use_environ (bool, optional): Use environment variable to determine main process.\nReturns:\n- bool: True if the current process is the main process, False otherwise.\n\n\n\nutils.distributed.reduce_and_broadcast(fn1, fn2)\nRun a callable ‘fn1’ on all ranks, gather the results, reduce them using ‘fn2’,\nand then broadcast the reduced result to all ranks.\nArgs:\n- fn1 (callable): A function that computes the value on each rank.\n- fn2 (callable): A reduction function that takes a list of values and returns a single value.\n- world_size (int, optional): Total number of processes in the current distributed setup.\nReturns:\n- The reduced and broadcasted value.\n\n\n\nutils.distributed.zero_first(is_main)\nruns the wrapped context so that rank 0 runs first before other ranks"
  },
  {
    "objectID": "docs/api/utils.schemas.datasets.html",
    "href": "docs/api/utils.schemas.datasets.html",
    "title": "utils.schemas.datasets",
    "section": "",
    "text": "utils.schemas.datasets\nPydantic models for datasets-related configuration\n\n\n\n\n\nName\nDescription\n\n\n\n\nDPODataset\nDPO configuration subset\n\n\nKTODataset\nKTO configuration subset\n\n\nPretrainingDataset\nPretraining dataset configuration subset\n\n\nSFTDataset\nSFT configuration subset\n\n\nStepwiseSupervisedDataset\nStepwise supervised dataset configuration subset\n\n\nUserDefinedDPOType\nUser defined typing for DPO\n\n\nUserDefinedKTOType\nUser defined typing for KTO\n\n\nUserDefinedPrompterType\nStructure for user defined prompt types\n\n\n\n\n\nutils.schemas.datasets.DPODataset()\nDPO configuration subset\n\n\n\nutils.schemas.datasets.KTODataset()\nKTO configuration subset\n\n\n\nutils.schemas.datasets.PretrainingDataset()\nPretraining dataset configuration subset\n\n\n\nutils.schemas.datasets.SFTDataset()\nSFT configuration subset\n\n\n\n\n\nName\nDescription\n\n\n\n\nhandle_legacy_message_fields\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\n\n\n\n\n\nutils.schemas.datasets.SFTDataset.handle_legacy_message_fields(data)\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\n\n\n\n\n\nutils.schemas.datasets.StepwiseSupervisedDataset()\nStepwise supervised dataset configuration subset\n\n\n\nutils.schemas.datasets.UserDefinedDPOType()\nUser defined typing for DPO\n\n\n\nutils.schemas.datasets.UserDefinedKTOType()\nUser defined typing for KTO\n\n\n\nutils.schemas.datasets.UserDefinedPrompterType()\nStructure for user defined prompt types"
  },
  {
    "objectID": "docs/api/utils.schemas.datasets.html#classes",
    "href": "docs/api/utils.schemas.datasets.html#classes",
    "title": "utils.schemas.datasets",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nDPODataset\nDPO configuration subset\n\n\nKTODataset\nKTO configuration subset\n\n\nPretrainingDataset\nPretraining dataset configuration subset\n\n\nSFTDataset\nSFT configuration subset\n\n\nStepwiseSupervisedDataset\nStepwise supervised dataset configuration subset\n\n\nUserDefinedDPOType\nUser defined typing for DPO\n\n\nUserDefinedKTOType\nUser defined typing for KTO\n\n\nUserDefinedPrompterType\nStructure for user defined prompt types\n\n\n\n\n\nutils.schemas.datasets.DPODataset()\nDPO configuration subset\n\n\n\nutils.schemas.datasets.KTODataset()\nKTO configuration subset\n\n\n\nutils.schemas.datasets.PretrainingDataset()\nPretraining dataset configuration subset\n\n\n\nutils.schemas.datasets.SFTDataset()\nSFT configuration subset\n\n\n\n\n\nName\nDescription\n\n\n\n\nhandle_legacy_message_fields\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\n\n\n\n\n\nutils.schemas.datasets.SFTDataset.handle_legacy_message_fields(data)\nHandle backwards compatibility between legacy message field mapping and new property mapping system.\n\n\n\n\n\nutils.schemas.datasets.StepwiseSupervisedDataset()\nStepwise supervised dataset configuration subset\n\n\n\nutils.schemas.datasets.UserDefinedDPOType()\nUser defined typing for DPO\n\n\n\nutils.schemas.datasets.UserDefinedKTOType()\nUser defined typing for KTO\n\n\n\nutils.schemas.datasets.UserDefinedPrompterType()\nStructure for user defined prompt types"
  },
  {
    "objectID": "docs/api/prompt_strategies.kto.user_defined.html",
    "href": "docs/api/prompt_strategies.kto.user_defined.html",
    "title": "prompt_strategies.kto.user_defined",
    "section": "",
    "text": "prompt_strategies.kto.user_defined\nprompt_strategies.kto.user_defined\nUser-defined KTO strategies"
  },
  {
    "objectID": "docs/api/monkeypatch.utils.html",
    "href": "docs/api/monkeypatch.utils.html",
    "title": "monkeypatch.utils",
    "section": "",
    "text": "monkeypatch.utils\nShared utils for the monkeypatches\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_cu_seqlens\ngenerate a cumulative sequence length mask for flash attention using attn mask\n\n\nget_cu_seqlens_from_pos_ids\ngenerate a cumulative sequence length mask for flash attention using pos ids\n\n\nmask_2d_to_4d\nExpands attention_mask from [bsz, seq_len] to [bsz, 1, tgt_seq_len, src_seq_len].\n\n\n\n\n\nmonkeypatch.utils.get_cu_seqlens(attn_mask)\ngenerate a cumulative sequence length mask for flash attention using attn mask\n\n\n\nmonkeypatch.utils.get_cu_seqlens_from_pos_ids(position_ids)\ngenerate a cumulative sequence length mask for flash attention using pos ids\n\n\n\nmonkeypatch.utils.mask_2d_to_4d(mask, dtype, tgt_len=None)\nExpands attention_mask from [bsz, seq_len] to [bsz, 1, tgt_seq_len, src_seq_len].\nThis expansion handles packed sequences so that sequences share the same attention mask integer value\nwhen they attend to each other within that sequence.\nThis expansion transforms the mask to lower triangular form to prevent future peeking."
  },
  {
    "objectID": "docs/api/monkeypatch.utils.html#functions",
    "href": "docs/api/monkeypatch.utils.html#functions",
    "title": "monkeypatch.utils",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nget_cu_seqlens\ngenerate a cumulative sequence length mask for flash attention using attn mask\n\n\nget_cu_seqlens_from_pos_ids\ngenerate a cumulative sequence length mask for flash attention using pos ids\n\n\nmask_2d_to_4d\nExpands attention_mask from [bsz, seq_len] to [bsz, 1, tgt_seq_len, src_seq_len].\n\n\n\n\n\nmonkeypatch.utils.get_cu_seqlens(attn_mask)\ngenerate a cumulative sequence length mask for flash attention using attn mask\n\n\n\nmonkeypatch.utils.get_cu_seqlens_from_pos_ids(position_ids)\ngenerate a cumulative sequence length mask for flash attention using pos ids\n\n\n\nmonkeypatch.utils.mask_2d_to_4d(mask, dtype, tgt_len=None)\nExpands attention_mask from [bsz, seq_len] to [bsz, 1, tgt_seq_len, src_seq_len].\nThis expansion handles packed sequences so that sequences share the same attention mask integer value\nwhen they attend to each other within that sequence.\nThis expansion transforms the mask to lower triangular form to prevent future peeking."
  },
  {
    "objectID": "docs/api/utils.schemas.integrations.html",
    "href": "docs/api/utils.schemas.integrations.html",
    "title": "utils.schemas.integrations",
    "section": "",
    "text": "utils.schemas.integrations\nPydantic models for Axolotl integrations\n\n\n\n\n\nName\nDescription\n\n\n\n\nCometConfig\nComet configuration subset\n\n\nGradioConfig\nGradio configuration subset\n\n\nLISAConfig\nLISA configuration subset\n\n\nMLFlowConfig\nMLFlow configuration subset\n\n\nRayConfig\nRay launcher configuration subset\n\n\nWandbConfig\nWandb configuration subset\n\n\n\n\n\nutils.schemas.integrations.CometConfig()\nComet configuration subset\n\n\n\nutils.schemas.integrations.GradioConfig()\nGradio configuration subset\n\n\n\nutils.schemas.integrations.LISAConfig()\nLISA configuration subset\n\n\n\nutils.schemas.integrations.MLFlowConfig()\nMLFlow configuration subset\n\n\n\nutils.schemas.integrations.RayConfig()\nRay launcher configuration subset\n\n\n\nutils.schemas.integrations.WandbConfig()\nWandb configuration subset"
  },
  {
    "objectID": "docs/api/utils.schemas.integrations.html#classes",
    "href": "docs/api/utils.schemas.integrations.html#classes",
    "title": "utils.schemas.integrations",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nCometConfig\nComet configuration subset\n\n\nGradioConfig\nGradio configuration subset\n\n\nLISAConfig\nLISA configuration subset\n\n\nMLFlowConfig\nMLFlow configuration subset\n\n\nRayConfig\nRay launcher configuration subset\n\n\nWandbConfig\nWandb configuration subset\n\n\n\n\n\nutils.schemas.integrations.CometConfig()\nComet configuration subset\n\n\n\nutils.schemas.integrations.GradioConfig()\nGradio configuration subset\n\n\n\nutils.schemas.integrations.LISAConfig()\nLISA configuration subset\n\n\n\nutils.schemas.integrations.MLFlowConfig()\nMLFlow configuration subset\n\n\n\nutils.schemas.integrations.RayConfig()\nRay launcher configuration subset\n\n\n\nutils.schemas.integrations.WandbConfig()\nWandb configuration subset"
  },
  {
    "objectID": "docs/api/core.trainers.mixins.rng_state_loader.html",
    "href": "docs/api/core.trainers.mixins.rng_state_loader.html",
    "title": "core.trainers.mixins.rng_state_loader",
    "section": "",
    "text": "core.trainers.mixins.rng_state_loader\nTemporary fix/override for bug in resume from checkpoint\nSee https://github.com/huggingface/transformers/pull/37162\nTODO: Remove when upstream added PR to release\n\n\n\n\n\nName\nDescription\n\n\n\n\nRngLoaderMixin\nmixin for method override to load RNG states from a checkpoint\n\n\n\n\n\ncore.trainers.mixins.rng_state_loader.RngLoaderMixin()\nmixin for method override to load RNG states from a checkpoint"
  },
  {
    "objectID": "docs/api/core.trainers.mixins.rng_state_loader.html#classes",
    "href": "docs/api/core.trainers.mixins.rng_state_loader.html#classes",
    "title": "core.trainers.mixins.rng_state_loader",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nRngLoaderMixin\nmixin for method override to load RNG states from a checkpoint\n\n\n\n\n\ncore.trainers.mixins.rng_state_loader.RngLoaderMixin()\nmixin for method override to load RNG states from a checkpoint"
  },
  {
    "objectID": "docs/api/integrations.lm_eval.args.html",
    "href": "docs/api/integrations.lm_eval.args.html",
    "title": "integrations.lm_eval.args",
    "section": "",
    "text": "integrations.lm_eval.args\nModule for handling lm eval harness input arguments.\n\n\n\n\n\nName\nDescription\n\n\n\n\nLMEvalArgs\nInput args for lm eval harness\n\n\n\n\n\nintegrations.lm_eval.args.LMEvalArgs()\nInput args for lm eval harness"
  },
  {
    "objectID": "docs/api/integrations.lm_eval.args.html#classes",
    "href": "docs/api/integrations.lm_eval.args.html#classes",
    "title": "integrations.lm_eval.args",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nLMEvalArgs\nInput args for lm eval harness\n\n\n\n\n\nintegrations.lm_eval.args.LMEvalArgs()\nInput args for lm eval harness"
  },
  {
    "objectID": "docs/api/loaders.processor.html",
    "href": "docs/api/loaders.processor.html",
    "title": "loaders.processor",
    "section": "",
    "text": "loaders.processor\nloaders.processor\nProcessor loading functionality for multi-modal models"
  },
  {
    "objectID": "docs/api/core.trainers.trl.html",
    "href": "docs/api/core.trainers.trl.html",
    "title": "core.trainers.trl",
    "section": "",
    "text": "core.trainers.trl\nModule for TRL PPO trainer\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlCPOTrainer\nExtend the base CPOTrainer for axolotl helpers\n\n\nAxolotlKTOTrainer\nExtend the base KTOTrainer for axolotl helpers\n\n\nAxolotlORPOTrainer\nExtend the base ORPOTrainer for axolotl helpers\n\n\nAxolotlPRMTrainer\nExtend the base trl.PRMTrainer for axolotl helpers\n\n\nAxolotlRewardTrainer\nExtend the base RewardTrainer for axolotl helpers\n\n\nTRLPPOTrainer\nWrapper for TRL PPO trainer to handle customizations\n\n\n\n\n\ncore.trainers.trl.AxolotlCPOTrainer(*args, **kwargs)\nExtend the base CPOTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlKTOTrainer(*args, **kwargs)\nExtend the base KTOTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlORPOTrainer(*args, **kwargs)\nExtend the base ORPOTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlPRMTrainer(*args, **kwargs)\nExtend the base trl.PRMTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlRewardTrainer(*args, **kwargs)\nExtend the base RewardTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.TRLPPOTrainer()\nWrapper for TRL PPO trainer to handle customizations"
  },
  {
    "objectID": "docs/api/core.trainers.trl.html#classes",
    "href": "docs/api/core.trainers.trl.html#classes",
    "title": "core.trainers.trl",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlCPOTrainer\nExtend the base CPOTrainer for axolotl helpers\n\n\nAxolotlKTOTrainer\nExtend the base KTOTrainer for axolotl helpers\n\n\nAxolotlORPOTrainer\nExtend the base ORPOTrainer for axolotl helpers\n\n\nAxolotlPRMTrainer\nExtend the base trl.PRMTrainer for axolotl helpers\n\n\nAxolotlRewardTrainer\nExtend the base RewardTrainer for axolotl helpers\n\n\nTRLPPOTrainer\nWrapper for TRL PPO trainer to handle customizations\n\n\n\n\n\ncore.trainers.trl.AxolotlCPOTrainer(*args, **kwargs)\nExtend the base CPOTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlKTOTrainer(*args, **kwargs)\nExtend the base KTOTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlORPOTrainer(*args, **kwargs)\nExtend the base ORPOTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlPRMTrainer(*args, **kwargs)\nExtend the base trl.PRMTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.AxolotlRewardTrainer(*args, **kwargs)\nExtend the base RewardTrainer for axolotl helpers\n\n\n\ncore.trainers.trl.TRLPPOTrainer()\nWrapper for TRL PPO trainer to handle customizations"
  },
  {
    "objectID": "docs/api/loaders.constants.html",
    "href": "docs/api/loaders.constants.html",
    "title": "loaders.constants",
    "section": "",
    "text": "loaders.constants\nloaders.constants\nShared constants for axolotl.loaders module"
  },
  {
    "objectID": "docs/api/monkeypatch.llama_attn_hijack_flash.html",
    "href": "docs/api/monkeypatch.llama_attn_hijack_flash.html",
    "title": "monkeypatch.llama_attn_hijack_flash",
    "section": "",
    "text": "monkeypatch.llama_attn_hijack_flash\nFlash attention monkey patch for llama model\n\n\n\n\n\nName\nDescription\n\n\n\n\nFusedAttention\nFused QKV Attention layer for incrementally improved training efficiency\n\n\nLlamaDecoderLayer\npatched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nmonkeypatch.llama_attn_hijack_flash.FusedAttention(config, q, k, v, o)\nFused QKV Attention layer for incrementally improved training efficiency\n\n\n\nmonkeypatch.llama_attn_hijack_flash.LlamaDecoderLayer()\npatched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nName\nDescription\n\n\n\n\nforward\n\n\n\n\n\n\nmonkeypatch.llama_attn_hijack_flash.LlamaDecoderLayer.forward(\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    padding_mask=None,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nhidden_states\ntorch.FloatTensor\ninput to the layer of shape (batch, seq_len, embed_dim)\nrequired\n\n\nattention_mask\ntorch.FloatTensor, optional\nattention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values.\nNone\n\n\noutput_attentions\nbool, optional\nWhether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.\nFalse\n\n\nuse_cache\nbool, optional\nIf set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).\nFalse\n\n\npast_key_value\nTuple(torch.FloatTensor), optional\ncached past key and value projection states\nNone\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nflashattn_forward\nInput shape: Batch x Time x Channel\n\n\nflashattn_forward_with_s2attn\nInput shape: Batch x Time x Channel\n\n\ngenerate_qkv\n\n\n\n\n\n\nmonkeypatch.llama_attn_hijack_flash.flashattn_forward(\n    self,\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    padding_mask=None,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\nInput shape: Batch x Time x Channel\nattention_mask: [bsz, q_len]\n\n\n\nmonkeypatch.llama_attn_hijack_flash.flashattn_forward_with_s2attn(\n    self,\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    padding_mask=None,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\nInput shape: Batch x Time x Channel\nFrom: https://github.com/dvlab-research/LongLoRA/blob/main/llama_attn_replace.py\nattention_mask: [bsz, q_len]\ncu_seqlens will be ignored if provided\nmax_seqlen will be ignored if provided\n\n\n\nmonkeypatch.llama_attn_hijack_flash.generate_qkv(\n    q,\n    k,\n    v,\n    query_padding_mask=None,\n    key_padding_mask=None,\n    kvpacked=False,\n    qkvpacked=False,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nq\n\n(batch_size, seqlen_q, nheads, d)\nrequired\n\n\nk\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nv\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nquery_padding_mask\n\n(batch_size, seqlen), bool\nNone\n\n\nkey_padding_mask\n\n(batch_size, seqlen), bool\nNone"
  },
  {
    "objectID": "docs/api/monkeypatch.llama_attn_hijack_flash.html#classes",
    "href": "docs/api/monkeypatch.llama_attn_hijack_flash.html#classes",
    "title": "monkeypatch.llama_attn_hijack_flash",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nFusedAttention\nFused QKV Attention layer for incrementally improved training efficiency\n\n\nLlamaDecoderLayer\npatched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nmonkeypatch.llama_attn_hijack_flash.FusedAttention(config, q, k, v, o)\nFused QKV Attention layer for incrementally improved training efficiency\n\n\n\nmonkeypatch.llama_attn_hijack_flash.LlamaDecoderLayer()\npatched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nName\nDescription\n\n\n\n\nforward\n\n\n\n\n\n\nmonkeypatch.llama_attn_hijack_flash.LlamaDecoderLayer.forward(\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    padding_mask=None,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nhidden_states\ntorch.FloatTensor\ninput to the layer of shape (batch, seq_len, embed_dim)\nrequired\n\n\nattention_mask\ntorch.FloatTensor, optional\nattention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values.\nNone\n\n\noutput_attentions\nbool, optional\nWhether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.\nFalse\n\n\nuse_cache\nbool, optional\nIf set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).\nFalse\n\n\npast_key_value\nTuple(torch.FloatTensor), optional\ncached past key and value projection states\nNone"
  },
  {
    "objectID": "docs/api/monkeypatch.llama_attn_hijack_flash.html#functions",
    "href": "docs/api/monkeypatch.llama_attn_hijack_flash.html#functions",
    "title": "monkeypatch.llama_attn_hijack_flash",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nflashattn_forward\nInput shape: Batch x Time x Channel\n\n\nflashattn_forward_with_s2attn\nInput shape: Batch x Time x Channel\n\n\ngenerate_qkv\n\n\n\n\n\n\nmonkeypatch.llama_attn_hijack_flash.flashattn_forward(\n    self,\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    padding_mask=None,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\nInput shape: Batch x Time x Channel\nattention_mask: [bsz, q_len]\n\n\n\nmonkeypatch.llama_attn_hijack_flash.flashattn_forward_with_s2attn(\n    self,\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    padding_mask=None,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\nInput shape: Batch x Time x Channel\nFrom: https://github.com/dvlab-research/LongLoRA/blob/main/llama_attn_replace.py\nattention_mask: [bsz, q_len]\ncu_seqlens will be ignored if provided\nmax_seqlen will be ignored if provided\n\n\n\nmonkeypatch.llama_attn_hijack_flash.generate_qkv(\n    q,\n    k,\n    v,\n    query_padding_mask=None,\n    key_padding_mask=None,\n    kvpacked=False,\n    qkvpacked=False,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nq\n\n(batch_size, seqlen_q, nheads, d)\nrequired\n\n\nk\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nv\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nquery_padding_mask\n\n(batch_size, seqlen), bool\nNone\n\n\nkey_padding_mask\n\n(batch_size, seqlen), bool\nNone"
  },
  {
    "objectID": "docs/api/cli.args.html",
    "href": "docs/api/cli.args.html",
    "title": "cli.args",
    "section": "",
    "text": "cli.args\nModule for axolotl CLI command arguments.\n\n\n\n\n\nName\nDescription\n\n\n\n\nEvaluateCliArgs\nDataclass with CLI arguments for axolotl evaluate command.\n\n\nInferenceCliArgs\nDataclass with CLI arguments for axolotl inference command.\n\n\nPreprocessCliArgs\nDataclass with CLI arguments for axolotl preprocess command.\n\n\nQuantizeCliArgs\nDataclass with CLI arguments for axolotl quantize command.\n\n\nTrainerCliArgs\nDataclass with CLI arguments for axolotl train command.\n\n\nVllmServeCliArgs\nDataclass with CLI arguments for axolotl vllm-serve command.\n\n\n\n\n\ncli.args.EvaluateCliArgs(\n    debug=False,\n    debug_text_only=False,\n    debug_num_examples=0,\n)\nDataclass with CLI arguments for axolotl evaluate command.\n\n\n\ncli.args.InferenceCliArgs(prompter=None)\nDataclass with CLI arguments for axolotl inference command.\n\n\n\ncli.args.PreprocessCliArgs(\n    debug=False,\n    debug_text_only=False,\n    debug_num_examples=1,\n    prompter=None,\n    download=True,\n    iterable=None,\n)\nDataclass with CLI arguments for axolotl preprocess command.\n\n\n\ncli.args.QuantizeCliArgs(\n    base_model=None,\n    weight_dtype=None,\n    activation_dtype=None,\n    quantize_embedding=None,\n    group_size=None,\n    output_dir=None,\n)\nDataclass with CLI arguments for axolotl quantize command.\n\n\n\ncli.args.TrainerCliArgs(\n    debug=False,\n    debug_text_only=False,\n    debug_num_examples=0,\n    prompter=None,\n    shard=False,\n    main_process_port=None,\n    num_processes=None,\n)\nDataclass with CLI arguments for axolotl train command.\n\n\n\ncli.args.VllmServeCliArgs(\n    tensor_parallel_size=None,\n    host=None,\n    port=None,\n    gpu_memory_utilization=None,\n    dtype=None,\n    max_model_len=None,\n    enable_prefix_caching=None,\n    serve_module=None,\n)\nDataclass with CLI arguments for axolotl vllm-serve command."
  },
  {
    "objectID": "docs/api/cli.args.html#classes",
    "href": "docs/api/cli.args.html#classes",
    "title": "cli.args",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nEvaluateCliArgs\nDataclass with CLI arguments for axolotl evaluate command.\n\n\nInferenceCliArgs\nDataclass with CLI arguments for axolotl inference command.\n\n\nPreprocessCliArgs\nDataclass with CLI arguments for axolotl preprocess command.\n\n\nQuantizeCliArgs\nDataclass with CLI arguments for axolotl quantize command.\n\n\nTrainerCliArgs\nDataclass with CLI arguments for axolotl train command.\n\n\nVllmServeCliArgs\nDataclass with CLI arguments for axolotl vllm-serve command.\n\n\n\n\n\ncli.args.EvaluateCliArgs(\n    debug=False,\n    debug_text_only=False,\n    debug_num_examples=0,\n)\nDataclass with CLI arguments for axolotl evaluate command.\n\n\n\ncli.args.InferenceCliArgs(prompter=None)\nDataclass with CLI arguments for axolotl inference command.\n\n\n\ncli.args.PreprocessCliArgs(\n    debug=False,\n    debug_text_only=False,\n    debug_num_examples=1,\n    prompter=None,\n    download=True,\n    iterable=None,\n)\nDataclass with CLI arguments for axolotl preprocess command.\n\n\n\ncli.args.QuantizeCliArgs(\n    base_model=None,\n    weight_dtype=None,\n    activation_dtype=None,\n    quantize_embedding=None,\n    group_size=None,\n    output_dir=None,\n)\nDataclass with CLI arguments for axolotl quantize command.\n\n\n\ncli.args.TrainerCliArgs(\n    debug=False,\n    debug_text_only=False,\n    debug_num_examples=0,\n    prompter=None,\n    shard=False,\n    main_process_port=None,\n    num_processes=None,\n)\nDataclass with CLI arguments for axolotl train command.\n\n\n\ncli.args.VllmServeCliArgs(\n    tensor_parallel_size=None,\n    host=None,\n    port=None,\n    gpu_memory_utilization=None,\n    dtype=None,\n    max_model_len=None,\n    enable_prefix_caching=None,\n    serve_module=None,\n)\nDataclass with CLI arguments for axolotl vllm-serve command."
  },
  {
    "objectID": "docs/api/core.builders.causal.html",
    "href": "docs/api/core.builders.causal.html",
    "title": "core.builders.causal",
    "section": "",
    "text": "core.builders.causal\nBuilder for causal trainers\n\n\n\n\n\nName\nDescription\n\n\n\n\nHFCausalTrainerBuilder\nBuild the HuggingFace training args/trainer for causal models and reward modeling\n\n\n\n\n\ncore.builders.causal.HFCausalTrainerBuilder(\n    cfg,\n    model,\n    tokenizer,\n    processor=None,\n)\nBuild the HuggingFace training args/trainer for causal models and reward modeling\nusing TRL."
  },
  {
    "objectID": "docs/api/core.builders.causal.html#classes",
    "href": "docs/api/core.builders.causal.html#classes",
    "title": "core.builders.causal",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nHFCausalTrainerBuilder\nBuild the HuggingFace training args/trainer for causal models and reward modeling\n\n\n\n\n\ncore.builders.causal.HFCausalTrainerBuilder(\n    cfg,\n    model,\n    tokenizer,\n    processor=None,\n)\nBuild the HuggingFace training args/trainer for causal models and reward modeling\nusing TRL."
  },
  {
    "objectID": "docs/api/prompt_strategies.input_output.html",
    "href": "docs/api/prompt_strategies.input_output.html",
    "title": "prompt_strategies.input_output",
    "section": "",
    "text": "prompt_strategies.input_output\nModule for plain input/output prompt pairs\n\n\n\n\n\nName\nDescription\n\n\n\n\nRawInputOutputPrompter\nprompter for raw i/o data\n\n\nRawInputOutputStrategy\nPrompt Strategy class for input/output pairs\n\n\n\n\n\nprompt_strategies.input_output.RawInputOutputPrompter()\nprompter for raw i/o data\n\n\n\nprompt_strategies.input_output.RawInputOutputStrategy(\n    *args,\n    eos_token=None,\n    **kwargs,\n)\nPrompt Strategy class for input/output pairs"
  },
  {
    "objectID": "docs/api/prompt_strategies.input_output.html#classes",
    "href": "docs/api/prompt_strategies.input_output.html#classes",
    "title": "prompt_strategies.input_output",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nRawInputOutputPrompter\nprompter for raw i/o data\n\n\nRawInputOutputStrategy\nPrompt Strategy class for input/output pairs\n\n\n\n\n\nprompt_strategies.input_output.RawInputOutputPrompter()\nprompter for raw i/o data\n\n\n\nprompt_strategies.input_output.RawInputOutputStrategy(\n    *args,\n    eos_token=None,\n    **kwargs,\n)\nPrompt Strategy class for input/output pairs"
  },
  {
    "objectID": "docs/api/core.chat.format.llama3x.html",
    "href": "docs/api/core.chat.format.llama3x.html",
    "title": "core.chat.format.llama3x",
    "section": "",
    "text": "core.chat.format.llama3x\ncore.chat.format.llama3x\nLlama 3.x chat formatting functions for MessageContents"
  },
  {
    "objectID": "docs/api/cli.train.html",
    "href": "docs/api/cli.train.html",
    "title": "cli.train",
    "section": "",
    "text": "cli.train\nCLI to run training on a model.\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_train.\n\n\ndo_train\nTrains a transformers model by first loading the dataset(s) specified in the\n\n\n\n\n\ncli.train.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_train.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.train.do_train(cfg, cli_args)\nTrains a transformers model by first loading the dataset(s) specified in the\naxolotl config, and then calling axolotl.train.train. Also runs the plugin\nmanager’s post_train_unload once training completes.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nTrainerCliArgs\nTraining-specific CLI arguments.\nrequired"
  },
  {
    "objectID": "docs/api/cli.train.html#functions",
    "href": "docs/api/cli.train.html#functions",
    "title": "cli.train",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_train.\n\n\ndo_train\nTrains a transformers model by first loading the dataset(s) specified in the\n\n\n\n\n\ncli.train.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_train.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.train.do_train(cfg, cli_args)\nTrains a transformers model by first loading the dataset(s) specified in the\naxolotl config, and then calling axolotl.train.train. Also runs the plugin\nmanager’s post_train_unload once training completes.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nTrainerCliArgs\nTraining-specific CLI arguments.\nrequired"
  },
  {
    "objectID": "docs/api/convert.html",
    "href": "docs/api/convert.html",
    "title": "convert",
    "section": "",
    "text": "convert\nModule containing File Reader, File Writer, Json Parser, and Jsonl Serializer classes\n\n\n\n\n\nName\nDescription\n\n\n\n\nFileReader\nReads a file and returns its contents as a string\n\n\nFileWriter\nWrites a string to a file\n\n\nJsonParser\nParses a string as JSON and returns the result\n\n\nJsonToJsonlConverter\nConverts a JSON file to JSONL\n\n\nJsonlSerializer\nSerializes a list of JSON objects into a JSONL string\n\n\nStdoutWriter\nWrites a string to stdout\n\n\n\n\n\nconvert.FileReader()\nReads a file and returns its contents as a string\n\n\n\nconvert.FileWriter(file_path)\nWrites a string to a file\n\n\n\nconvert.JsonParser()\nParses a string as JSON and returns the result\n\n\n\nconvert.JsonToJsonlConverter(\n    file_reader,\n    file_writer,\n    json_parser,\n    jsonl_serializer,\n)\nConverts a JSON file to JSONL\n\n\n\nconvert.JsonlSerializer()\nSerializes a list of JSON objects into a JSONL string\n\n\n\nconvert.StdoutWriter()\nWrites a string to stdout"
  },
  {
    "objectID": "docs/api/convert.html#classes",
    "href": "docs/api/convert.html#classes",
    "title": "convert",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nFileReader\nReads a file and returns its contents as a string\n\n\nFileWriter\nWrites a string to a file\n\n\nJsonParser\nParses a string as JSON and returns the result\n\n\nJsonToJsonlConverter\nConverts a JSON file to JSONL\n\n\nJsonlSerializer\nSerializes a list of JSON objects into a JSONL string\n\n\nStdoutWriter\nWrites a string to stdout\n\n\n\n\n\nconvert.FileReader()\nReads a file and returns its contents as a string\n\n\n\nconvert.FileWriter(file_path)\nWrites a string to a file\n\n\n\nconvert.JsonParser()\nParses a string as JSON and returns the result\n\n\n\nconvert.JsonToJsonlConverter(\n    file_reader,\n    file_writer,\n    json_parser,\n    jsonl_serializer,\n)\nConverts a JSON file to JSONL\n\n\n\nconvert.JsonlSerializer()\nSerializes a list of JSON objects into a JSONL string\n\n\n\nconvert.StdoutWriter()\nWrites a string to stdout"
  },
  {
    "objectID": "docs/api/utils.optimizers.adopt.html",
    "href": "docs/api/utils.optimizers.adopt.html",
    "title": "utils.optimizers.adopt",
    "section": "",
    "text": "utils.optimizers.adopt\nCopied from https://github.com/iShohei220/adopt\nADOPT: Modified Adam Can Converge with Any β2 with the Optimal Rate (2024)\nTaniguchi, Shohei and Harada, Keno and Minegishi, Gouki and Oshima, Yuta and Jeong, Seong Cheol and Nagahara, Go and Iiyama, Tomoshi and Suzuki, Masahiro and Iwasawa, Yusuke and Matsuo, Yutaka\n\n\n\n\n\nName\nDescription\n\n\n\n\nadopt\nFunctional API that performs ADOPT algorithm computation.\n\n\n\n\n\nutils.optimizers.adopt.adopt(\n    params,\n    grads,\n    exp_avgs,\n    exp_avg_sqs,\n    state_steps,\n    foreach=None,\n    capturable=False,\n    differentiable=False,\n    fused=None,\n    grad_scale=None,\n    found_inf=None,\n    has_complex=False,\n    *,\n    beta1,\n    beta2,\n    lr,\n    clip_lambda,\n    weight_decay,\n    decouple,\n    eps,\n    maximize,\n)\nFunctional API that performs ADOPT algorithm computation."
  },
  {
    "objectID": "docs/api/utils.optimizers.adopt.html#functions",
    "href": "docs/api/utils.optimizers.adopt.html#functions",
    "title": "utils.optimizers.adopt",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nadopt\nFunctional API that performs ADOPT algorithm computation.\n\n\n\n\n\nutils.optimizers.adopt.adopt(\n    params,\n    grads,\n    exp_avgs,\n    exp_avg_sqs,\n    state_steps,\n    foreach=None,\n    capturable=False,\n    differentiable=False,\n    fused=None,\n    grad_scale=None,\n    found_inf=None,\n    has_complex=False,\n    *,\n    beta1,\n    beta2,\n    lr,\n    clip_lambda,\n    weight_decay,\n    decouple,\n    eps,\n    maximize,\n)\nFunctional API that performs ADOPT algorithm computation."
  },
  {
    "objectID": "docs/api/prompt_strategies.pygmalion.html",
    "href": "docs/api/prompt_strategies.pygmalion.html",
    "title": "prompt_strategies.pygmalion",
    "section": "",
    "text": "prompt_strategies.pygmalion\nModule containing the PygmalionPromptTokenizingStrategy and PygmalionPrompter class\n\n\n\n\n\nName\nDescription\n\n\n\n\nPygmalionPromptTokenizingStrategy\nTokenizing strategy for Pygmalion.\n\n\nPygmalionPrompter\nPrompter for Pygmalion.\n\n\n\n\n\nprompt_strategies.pygmalion.PygmalionPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    *args,\n    **kwargs,\n)\nTokenizing strategy for Pygmalion.\n\n\n\nprompt_strategies.pygmalion.PygmalionPrompter(*args, **kwargs)\nPrompter for Pygmalion."
  },
  {
    "objectID": "docs/api/prompt_strategies.pygmalion.html#classes",
    "href": "docs/api/prompt_strategies.pygmalion.html#classes",
    "title": "prompt_strategies.pygmalion",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nPygmalionPromptTokenizingStrategy\nTokenizing strategy for Pygmalion.\n\n\nPygmalionPrompter\nPrompter for Pygmalion.\n\n\n\n\n\nprompt_strategies.pygmalion.PygmalionPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    *args,\n    **kwargs,\n)\nTokenizing strategy for Pygmalion.\n\n\n\nprompt_strategies.pygmalion.PygmalionPrompter(*args, **kwargs)\nPrompter for Pygmalion."
  },
  {
    "objectID": "docs/api/integrations.kd.trainer.html",
    "href": "docs/api/integrations.kd.trainer.html",
    "title": "integrations.kd.trainer",
    "section": "",
    "text": "integrations.kd.trainer\nKD trainer\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlKDTrainer\nCustom trainer subclass for Knowledge Distillation (KD)\n\n\n\n\n\nintegrations.kd.trainer.AxolotlKDTrainer(\n    *_args,\n    bench_data_collator=None,\n    eval_data_collator=None,\n    dataset_tags=None,\n    **kwargs,\n)\nCustom trainer subclass for Knowledge Distillation (KD)\n\n\n\n\n\nName\nDescription\n\n\n\n\ncompute_loss\nHow the loss is computed by Trainer. By default, all models return the loss in the first element.\n\n\n\n\n\nintegrations.kd.trainer.AxolotlKDTrainer.compute_loss(\n    model,\n    inputs,\n    return_outputs=False,\n    num_items_in_batch=None,\n)\nHow the loss is computed by Trainer. By default, all models return the loss in the first element.\nSubclass and override for custom behavior."
  },
  {
    "objectID": "docs/api/integrations.kd.trainer.html#classes",
    "href": "docs/api/integrations.kd.trainer.html#classes",
    "title": "integrations.kd.trainer",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlKDTrainer\nCustom trainer subclass for Knowledge Distillation (KD)\n\n\n\n\n\nintegrations.kd.trainer.AxolotlKDTrainer(\n    *_args,\n    bench_data_collator=None,\n    eval_data_collator=None,\n    dataset_tags=None,\n    **kwargs,\n)\nCustom trainer subclass for Knowledge Distillation (KD)\n\n\n\n\n\nName\nDescription\n\n\n\n\ncompute_loss\nHow the loss is computed by Trainer. By default, all models return the loss in the first element.\n\n\n\n\n\nintegrations.kd.trainer.AxolotlKDTrainer.compute_loss(\n    model,\n    inputs,\n    return_outputs=False,\n    num_items_in_batch=None,\n)\nHow the loss is computed by Trainer. By default, all models return the loss in the first element.\nSubclass and override for custom behavior."
  },
  {
    "objectID": "docs/api/utils.callbacks.qat.html",
    "href": "docs/api/utils.callbacks.qat.html",
    "title": "utils.callbacks.qat",
    "section": "",
    "text": "utils.callbacks.qat\nQAT Callback for HF Causal Trainer\n\n\n\n\n\nName\nDescription\n\n\n\n\nQATCallback\nCallback to toggle fake quantization for the model.\n\n\n\n\n\nutils.callbacks.qat.QATCallback(cfg)\nCallback to toggle fake quantization for the model.\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\ntoggle_fake_quant\nToggle fake quantization for any fake quantized linear or embedding layers in the model.\n\n\n\n\n\nutils.callbacks.qat.toggle_fake_quant(mod, enable)\nToggle fake quantization for any fake quantized linear or embedding layers in the model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmod\nnn.Module\nThe module to toggle fake quantization for.\nrequired\n\n\nenable\nbool\nWhether to enable or disable fake quantization.\nrequired"
  },
  {
    "objectID": "docs/api/utils.callbacks.qat.html#classes",
    "href": "docs/api/utils.callbacks.qat.html#classes",
    "title": "utils.callbacks.qat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nQATCallback\nCallback to toggle fake quantization for the model.\n\n\n\n\n\nutils.callbacks.qat.QATCallback(cfg)\nCallback to toggle fake quantization for the model."
  },
  {
    "objectID": "docs/api/utils.callbacks.qat.html#functions",
    "href": "docs/api/utils.callbacks.qat.html#functions",
    "title": "utils.callbacks.qat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ntoggle_fake_quant\nToggle fake quantization for any fake quantized linear or embedding layers in the model.\n\n\n\n\n\nutils.callbacks.qat.toggle_fake_quant(mod, enable)\nToggle fake quantization for any fake quantized linear or embedding layers in the model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmod\nnn.Module\nThe module to toggle fake quantization for.\nrequired\n\n\nenable\nbool\nWhether to enable or disable fake quantization.\nrequired"
  },
  {
    "objectID": "docs/api/utils.callbacks.lisa.html",
    "href": "docs/api/utils.callbacks.lisa.html",
    "title": "utils.callbacks.lisa",
    "section": "",
    "text": "utils.callbacks.lisa\nutils.callbacks.lisa\nmodule for LISA\nAdapted from https://github.com/OptimalScale/LMFlow/pull/701 for HF transformers & Axolotl\nArxiv: https://arxiv.org/abs/2403.17919\nLicense: Apache 2.0"
  },
  {
    "objectID": "docs/api/loaders.tokenizer.html",
    "href": "docs/api/loaders.tokenizer.html",
    "title": "loaders.tokenizer",
    "section": "",
    "text": "loaders.tokenizer\nTokenizer loading functionality and associated utils\n\n\n\n\n\nName\nDescription\n\n\n\n\nload_tokenizer\nLoad and configure the tokenizer based on the provided config.\n\n\nmodify_tokenizer_files\nModify tokenizer files to replace added_tokens strings, save to output directory,\n\n\n\n\n\nloaders.tokenizer.load_tokenizer(cfg)\nLoad and configure the tokenizer based on the provided config.\n\n\n\nloaders.tokenizer.modify_tokenizer_files(\n    tokenizer_path,\n    token_mappings,\n    output_dir,\n)\nModify tokenizer files to replace added_tokens strings, save to output directory,\nand return the path to the modified tokenizer.\nThis only works with reserved tokens that were added to the tokenizer, not tokens\nalready part of the vocab.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer_path\nstr\nPath or name of the original tokenizer\nrequired\n\n\ntoken_mappings\ndict[int, str]\nDict mapping {token_id (int): new_token_string}\nrequired\n\n\noutput_dir\nstr\nDirectory to save the modified tokenizer\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr\nPath to the modified tokenizer directory\n\n\n\nRef: https://github.com/huggingface/transformers/issues/27974#issuecomment-1854188941"
  },
  {
    "objectID": "docs/api/loaders.tokenizer.html#functions",
    "href": "docs/api/loaders.tokenizer.html#functions",
    "title": "loaders.tokenizer",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nload_tokenizer\nLoad and configure the tokenizer based on the provided config.\n\n\nmodify_tokenizer_files\nModify tokenizer files to replace added_tokens strings, save to output directory,\n\n\n\n\n\nloaders.tokenizer.load_tokenizer(cfg)\nLoad and configure the tokenizer based on the provided config.\n\n\n\nloaders.tokenizer.modify_tokenizer_files(\n    tokenizer_path,\n    token_mappings,\n    output_dir,\n)\nModify tokenizer files to replace added_tokens strings, save to output directory,\nand return the path to the modified tokenizer.\nThis only works with reserved tokens that were added to the tokenizer, not tokens\nalready part of the vocab.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer_path\nstr\nPath or name of the original tokenizer\nrequired\n\n\ntoken_mappings\ndict[int, str]\nDict mapping {token_id (int): new_token_string}\nrequired\n\n\noutput_dir\nstr\nDirectory to save the modified tokenizer\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr\nPath to the modified tokenizer directory\n\n\n\nRef: https://github.com/huggingface/transformers/issues/27974#issuecomment-1854188941"
  },
  {
    "objectID": "docs/api/cli.main.html",
    "href": "docs/api/cli.main.html",
    "title": "cli.main",
    "section": "",
    "text": "cli.main\nClick CLI definitions for various axolotl commands.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncli\nAxolotl CLI - Train and fine-tune large language models\n\n\nevaluate\nEvaluate a model.\n\n\nfetch\nFetch example configs or other resources.\n\n\ninference\nRun inference with a trained model.\n\n\nmerge_lora\nMerge trained LoRA adapters into a base model.\n\n\nmerge_sharded_fsdp_weights\nMerge sharded FSDP model weights.\n\n\npreprocess\nPreprocess datasets before training.\n\n\ntrain\nTrain or fine-tune a model.\n\n\n\n\n\ncli.main.cli()\nAxolotl CLI - Train and fine-tune large language models\n\n\n\ncli.main.evaluate(config, accelerate, **kwargs)\nEvaluate a model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.fetch(directory, dest)\nFetch example configs or other resources.\nAvailable directories:\n- examples: Example configuration files\n- deepspeed_configs: DeepSpeed configuration files\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndirectory\nstr\nOne of examples, deepspeed_configs.\nrequired\n\n\ndest\nOptional[str]\nOptional destination directory.\nrequired\n\n\n\n\n\n\n\ncli.main.inference(config, accelerate, gradio, **kwargs)\nRun inference with a trained model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\ngradio\nbool\nWhether to use Gradio browser interface or command line for inference.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.merge_lora(config, **kwargs)\nMerge trained LoRA adapters into a base model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.merge_sharded_fsdp_weights(config, accelerate, **kwargs)\nMerge sharded FSDP model weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.preprocess(config, cloud=None, **kwargs)\nPreprocess datasets before training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\ncloud\nOptional[str]\nPath to a cloud accelerator configuration file.\nNone\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.train(config, accelerate, cloud=None, sweep=None, **kwargs)\nTrain or fine-tune a model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\ncloud\nOptional[str]\nPath to a cloud accelerator configuration file\nNone\n\n\nsweep\nOptional[str]\nPath to YAML config for sweeping hyperparameters.\nNone\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}"
  },
  {
    "objectID": "docs/api/cli.main.html#functions",
    "href": "docs/api/cli.main.html#functions",
    "title": "cli.main",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ncli\nAxolotl CLI - Train and fine-tune large language models\n\n\nevaluate\nEvaluate a model.\n\n\nfetch\nFetch example configs or other resources.\n\n\ninference\nRun inference with a trained model.\n\n\nmerge_lora\nMerge trained LoRA adapters into a base model.\n\n\nmerge_sharded_fsdp_weights\nMerge sharded FSDP model weights.\n\n\npreprocess\nPreprocess datasets before training.\n\n\ntrain\nTrain or fine-tune a model.\n\n\n\n\n\ncli.main.cli()\nAxolotl CLI - Train and fine-tune large language models\n\n\n\ncli.main.evaluate(config, accelerate, **kwargs)\nEvaluate a model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.fetch(directory, dest)\nFetch example configs or other resources.\nAvailable directories:\n- examples: Example configuration files\n- deepspeed_configs: DeepSpeed configuration files\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ndirectory\nstr\nOne of examples, deepspeed_configs.\nrequired\n\n\ndest\nOptional[str]\nOptional destination directory.\nrequired\n\n\n\n\n\n\n\ncli.main.inference(config, accelerate, gradio, **kwargs)\nRun inference with a trained model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\ngradio\nbool\nWhether to use Gradio browser interface or command line for inference.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.merge_lora(config, **kwargs)\nMerge trained LoRA adapters into a base model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.merge_sharded_fsdp_weights(config, accelerate, **kwargs)\nMerge sharded FSDP model weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.preprocess(config, cloud=None, **kwargs)\nPreprocess datasets before training.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\ncloud\nOptional[str]\nPath to a cloud accelerator configuration file.\nNone\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}\n\n\n\n\n\n\n\ncli.main.train(config, accelerate, cloud=None, sweep=None, **kwargs)\nTrain or fine-tune a model.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr\nPath to axolotl config YAML file.\nrequired\n\n\naccelerate\nbool\nWhether to use accelerate launcher.\nrequired\n\n\ncloud\nOptional[str]\nPath to a cloud accelerator configuration file\nNone\n\n\nsweep\nOptional[str]\nPath to YAML config for sweeping hyperparameters.\nNone\n\n\nkwargs\n\nAdditional keyword arguments which correspond to CLI args or axolotl config options.\n{}"
  },
  {
    "objectID": "docs/api/monkeypatch.data.batch_dataset_fetcher.html",
    "href": "docs/api/monkeypatch.data.batch_dataset_fetcher.html",
    "title": "monkeypatch.data.batch_dataset_fetcher",
    "section": "",
    "text": "monkeypatch.data.batch_dataset_fetcher\nmonkeypatch.data.batch_dataset_fetcher\nmonkey patches for the dataset fetcher to handle batches of packed indexes"
  },
  {
    "objectID": "docs/api/cli.evaluate.html",
    "href": "docs/api/cli.evaluate.html",
    "title": "cli.evaluate",
    "section": "",
    "text": "cli.evaluate\nCLI to run evaluation on a model.\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_evaluate.\n\n\ndo_evaluate\nEvaluates a transformers model by first loading the dataset(s) specified in the\n\n\n\n\n\ncli.evaluate.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_evaluate.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.evaluate.do_evaluate(cfg, cli_args)\nEvaluates a transformers model by first loading the dataset(s) specified in the\naxolotl config, and then calling axolotl.evaluate.evaluate, which computes\nevaluation metrics on the given dataset(s) and writes them to disk.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nTrainerCliArgs\nCLI arguments.\nrequired"
  },
  {
    "objectID": "docs/api/cli.evaluate.html#functions",
    "href": "docs/api/cli.evaluate.html#functions",
    "title": "cli.evaluate",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_evaluate.\n\n\ndo_evaluate\nEvaluates a transformers model by first loading the dataset(s) specified in the\n\n\n\n\n\ncli.evaluate.do_cli(config=Path('examples/'), **kwargs)\nParses axolotl config, CLI args, and calls do_evaluate.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.evaluate.do_evaluate(cfg, cli_args)\nEvaluates a transformers model by first loading the dataset(s) specified in the\naxolotl config, and then calling axolotl.evaluate.evaluate, which computes\nevaluation metrics on the given dataset(s) and writes them to disk.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nTrainerCliArgs\nCLI arguments.\nrequired"
  },
  {
    "objectID": "docs/api/utils.bench.html",
    "href": "docs/api/utils.bench.html",
    "title": "utils.bench",
    "section": "",
    "text": "utils.bench\nBenchmarking and measurement utilities\n\n\n\n\n\nName\nDescription\n\n\n\n\ncheck_cuda_device\nwraps a function and returns the default value instead of running the\n\n\n\n\n\nutils.bench.check_cuda_device(default_value)\nwraps a function and returns the default value instead of running the\nwrapped function if cuda isn’t available or the device is auto\n:param default_value:\n:return:"
  },
  {
    "objectID": "docs/api/utils.bench.html#functions",
    "href": "docs/api/utils.bench.html#functions",
    "title": "utils.bench",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ncheck_cuda_device\nwraps a function and returns the default value instead of running the\n\n\n\n\n\nutils.bench.check_cuda_device(default_value)\nwraps a function and returns the default value instead of running the\nwrapped function if cuda isn’t available or the device is auto\n:param default_value:\n:return:"
  },
  {
    "objectID": "docs/api/monkeypatch.transformers_fa_utils.html",
    "href": "docs/api/monkeypatch.transformers_fa_utils.html",
    "title": "monkeypatch.transformers_fa_utils",
    "section": "",
    "text": "monkeypatch.transformers_fa_utils\nsee https://github.com/huggingface/transformers/pull/35834\n\n\n\n\n\nName\nDescription\n\n\n\n\nfixed_fa_peft_integration_check\nPEFT usually casts the layer norms in float32 for training stability reasons\n\n\n\n\n\nmonkeypatch.transformers_fa_utils.fixed_fa_peft_integration_check(\n    query,\n    key,\n    value,\n    target_dtype=None,\n    preferred_dtype=None,\n)\nPEFT usually casts the layer norms in float32 for training stability reasons\ntherefore the input hidden states gets silently casted in float32. Hence, we need\ncast them back in float16 / bfloat16 just to be sure everything works as expected.\nThis might slowdown training & inference so it is recommended to not cast the LayerNorms!\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nquery\ntorch.Tensor\nInput query states to be passed to Flash Attention API\nrequired\n\n\nkey\ntorch.Tensor\nInput key states to be passed to Flash Attention API\nrequired\n\n\nvalue\ntorch.Tensor\nInput value states to be passed to Flash Attention API\nrequired\n\n\ntarget_dtype\ntorch.dtype, optional\nThe dtype to convert the attention tensors to. Conversion can be ignored by not providing the target dtype.\nNone\n\n\npreferred_dtype\ntorch.dtype, optional\nThe preferred dtype to convert the attention tensors to regardless of the target dtype.\nNone"
  },
  {
    "objectID": "docs/api/monkeypatch.transformers_fa_utils.html#functions",
    "href": "docs/api/monkeypatch.transformers_fa_utils.html#functions",
    "title": "monkeypatch.transformers_fa_utils",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nfixed_fa_peft_integration_check\nPEFT usually casts the layer norms in float32 for training stability reasons\n\n\n\n\n\nmonkeypatch.transformers_fa_utils.fixed_fa_peft_integration_check(\n    query,\n    key,\n    value,\n    target_dtype=None,\n    preferred_dtype=None,\n)\nPEFT usually casts the layer norms in float32 for training stability reasons\ntherefore the input hidden states gets silently casted in float32. Hence, we need\ncast them back in float16 / bfloat16 just to be sure everything works as expected.\nThis might slowdown training & inference so it is recommended to not cast the LayerNorms!\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nquery\ntorch.Tensor\nInput query states to be passed to Flash Attention API\nrequired\n\n\nkey\ntorch.Tensor\nInput key states to be passed to Flash Attention API\nrequired\n\n\nvalue\ntorch.Tensor\nInput value states to be passed to Flash Attention API\nrequired\n\n\ntarget_dtype\ntorch.dtype, optional\nThe dtype to convert the attention tensors to. Conversion can be ignored by not providing the target dtype.\nNone\n\n\npreferred_dtype\ntorch.dtype, optional\nThe preferred dtype to convert the attention tensors to regardless of the target dtype.\nNone"
  },
  {
    "objectID": "docs/api/monkeypatch.llama_patch_multipack.html",
    "href": "docs/api/monkeypatch.llama_patch_multipack.html",
    "title": "monkeypatch.llama_patch_multipack",
    "section": "",
    "text": "monkeypatch.llama_patch_multipack\nmonkeypatch.llama_patch_multipack\nPatched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention"
  },
  {
    "objectID": "docs/api/monkeypatch.mistral_attn_hijack_flash.html",
    "href": "docs/api/monkeypatch.mistral_attn_hijack_flash.html",
    "title": "monkeypatch.mistral_attn_hijack_flash",
    "section": "",
    "text": "monkeypatch.mistral_attn_hijack_flash\nFlash attention monkey patch for mistral model\n\n\n\n\n\nName\nDescription\n\n\n\n\nMistralDecoderLayer\npatched version of MistralDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nmonkeypatch.mistral_attn_hijack_flash.MistralDecoderLayer()\npatched version of MistralDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nName\nDescription\n\n\n\n\nforward\n\n\n\n\n\n\nmonkeypatch.mistral_attn_hijack_flash.MistralDecoderLayer.forward(\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nhidden_states\ntorch.FloatTensor\ninput to the layer of shape (batch, seq_len, embed_dim)\nrequired\n\n\nattention_mask\ntorch.FloatTensor, optional\nattention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values.\nNone\n\n\noutput_attentions\nbool, optional\nWhether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.\nFalse\n\n\nuse_cache\nbool, optional\nIf set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).\nFalse\n\n\npast_key_value\nTuple(torch.FloatTensor), optional\ncached past key and value projection states\nNone\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\ngenerate_qkv\n\n\n\n\n\n\nmonkeypatch.mistral_attn_hijack_flash.generate_qkv(\n    q,\n    k,\n    v,\n    query_padding_mask=None,\n    key_padding_mask=None,\n    kvpacked=False,\n    qkvpacked=False,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nq\n\n(batch_size, seqlen_q, nheads, d)\nrequired\n\n\nk\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nv\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nquery_padding_mask\n\n(batch_size, seqlen), bool\nNone\n\n\nkey_padding_mask\n\n(batch_size, seqlen), bool\nNone"
  },
  {
    "objectID": "docs/api/monkeypatch.mistral_attn_hijack_flash.html#classes",
    "href": "docs/api/monkeypatch.mistral_attn_hijack_flash.html#classes",
    "title": "monkeypatch.mistral_attn_hijack_flash",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMistralDecoderLayer\npatched version of MistralDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nmonkeypatch.mistral_attn_hijack_flash.MistralDecoderLayer()\npatched version of MistralDecoderLayer to pass through the precalculated cu_seqlens\n\n\n\n\n\nName\nDescription\n\n\n\n\nforward\n\n\n\n\n\n\nmonkeypatch.mistral_attn_hijack_flash.MistralDecoderLayer.forward(\n    hidden_states,\n    attention_mask=None,\n    position_ids=None,\n    past_key_value=None,\n    output_attentions=False,\n    use_cache=False,\n    cu_seqlens=None,\n    max_seqlen=None,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nhidden_states\ntorch.FloatTensor\ninput to the layer of shape (batch, seq_len, embed_dim)\nrequired\n\n\nattention_mask\ntorch.FloatTensor, optional\nattention mask of size (batch, 1, tgt_len, src_len) where padding elements are indicated by very large negative values.\nNone\n\n\noutput_attentions\nbool, optional\nWhether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.\nFalse\n\n\nuse_cache\nbool, optional\nIf set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).\nFalse\n\n\npast_key_value\nTuple(torch.FloatTensor), optional\ncached past key and value projection states\nNone"
  },
  {
    "objectID": "docs/api/monkeypatch.mistral_attn_hijack_flash.html#functions",
    "href": "docs/api/monkeypatch.mistral_attn_hijack_flash.html#functions",
    "title": "monkeypatch.mistral_attn_hijack_flash",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ngenerate_qkv\n\n\n\n\n\n\nmonkeypatch.mistral_attn_hijack_flash.generate_qkv(\n    q,\n    k,\n    v,\n    query_padding_mask=None,\n    key_padding_mask=None,\n    kvpacked=False,\n    qkvpacked=False,\n)\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nq\n\n(batch_size, seqlen_q, nheads, d)\nrequired\n\n\nk\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nv\n\n(batch_size, seqlen_k, nheads_k, d)\nrequired\n\n\nquery_padding_mask\n\n(batch_size, seqlen), bool\nNone\n\n\nkey_padding_mask\n\n(batch_size, seqlen), bool\nNone"
  },
  {
    "objectID": "docs/api/utils.ctx_managers.sequence_parallel.html",
    "href": "docs/api/utils.ctx_managers.sequence_parallel.html",
    "title": "utils.ctx_managers.sequence_parallel",
    "section": "",
    "text": "utils.ctx_managers.sequence_parallel\nModule for Axolotl trainer sequence parallelism manager and utilities\n\n\n\n\n\nName\nDescription\n\n\n\n\nAllGatherWithGrad\nCustom autograd function for all-gather to preserve gradients.\n\n\nSequenceParallelContextManager\nContext manager for sequence parallelism operations.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad()\nCustom autograd function for all-gather to preserve gradients.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass for all-gather operation.\n\n\nforward\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(\n    ctx,\n    grad_output,\n)\nBackward pass for all-gather operation.\nExtracts the gradient slice corresponding to this rank’s original input\nfrom the full gradient tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient from subsequent layers with respect to the concatenated output tensor.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None]\nTuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(\n    ctx,\n    input_tensor,\n    group,\n)\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ninput_tensor\ntorch.Tensor\nTensor from model output with sequence dimension.\nrequired\n\n\ngroup\ndist.ProcessGroup\ntorch.distributed process group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.\n\n\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.SequenceParallelContextManager(\n    models,\n    sequence_parallel_degree,\n    gradient_accumulation_steps,\n    ring_attn_func,\n    heads_k_stride,\n)\nContext manager for sequence parallelism operations.\nThis class provides a context that will automatically apply sequence parallelism\nduring model forward passes using a pre-forward hook, and gather outputs from\nacross the sequence parallelism group using a post-forward hook.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodels\nlist[nn.Module]\nList of models to apply sequence parallelism to pre- and post- forward hooks.\nrequired\n\n\nsequence_parallel_degree\nint\nNumber of processes to split sequences over.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused.\nrequired\n\n\nheads_k_stride\nint | None\nSequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.\nrequired\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_sequence_parallelism\nApply sequence parallelism slicing to a batch.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.apply_sequence_parallelism(\n    batch,\n    local_rank,\n    local_world_size,\n    gradient_accumulation_steps,\n    ring_attn_func,\n)\nApply sequence parallelism slicing to a batch.\nSpecial handling is implemented for integer logits_to_keep, which indicates\nto only keep the last N tokens in the sequence during generation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbatch\ndict[str, torch.Tensor]\nBatch dictionary (e.g., input_ids, attention_mask, etc.).\nrequired\n\n\nlocal_rank\nint\nLocal rank in the sequence parallel group.\nrequired\n\n\nlocal_world_size\nint\nWorld size of the sequence parallel group.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused, but related to above TODO.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[dict[str, torch.Tensor], int, int]\ntuple of: - Batch dictionary with sliced tensors. - The original sequence length before padding. - The number of padding tokens added."
  },
  {
    "objectID": "docs/api/utils.ctx_managers.sequence_parallel.html#classes",
    "href": "docs/api/utils.ctx_managers.sequence_parallel.html#classes",
    "title": "utils.ctx_managers.sequence_parallel",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAllGatherWithGrad\nCustom autograd function for all-gather to preserve gradients.\n\n\nSequenceParallelContextManager\nContext manager for sequence parallelism operations.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad()\nCustom autograd function for all-gather to preserve gradients.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass for all-gather operation.\n\n\nforward\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(\n    ctx,\n    grad_output,\n)\nBackward pass for all-gather operation.\nExtracts the gradient slice corresponding to this rank’s original input\nfrom the full gradient tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient from subsequent layers with respect to the concatenated output tensor.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None]\nTuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(\n    ctx,\n    input_tensor,\n    group,\n)\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ninput_tensor\ntorch.Tensor\nTensor from model output with sequence dimension.\nrequired\n\n\ngroup\ndist.ProcessGroup\ntorch.distributed process group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.\n\n\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.SequenceParallelContextManager(\n    models,\n    sequence_parallel_degree,\n    gradient_accumulation_steps,\n    ring_attn_func,\n    heads_k_stride,\n)\nContext manager for sequence parallelism operations.\nThis class provides a context that will automatically apply sequence parallelism\nduring model forward passes using a pre-forward hook, and gather outputs from\nacross the sequence parallelism group using a post-forward hook.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodels\nlist[nn.Module]\nList of models to apply sequence parallelism to pre- and post- forward hooks.\nrequired\n\n\nsequence_parallel_degree\nint\nNumber of processes to split sequences over.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused.\nrequired\n\n\nheads_k_stride\nint | None\nSequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.\nrequired"
  },
  {
    "objectID": "docs/api/utils.ctx_managers.sequence_parallel.html#functions",
    "href": "docs/api/utils.ctx_managers.sequence_parallel.html#functions",
    "title": "utils.ctx_managers.sequence_parallel",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\napply_sequence_parallelism\nApply sequence parallelism slicing to a batch.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.apply_sequence_parallelism(\n    batch,\n    local_rank,\n    local_world_size,\n    gradient_accumulation_steps,\n    ring_attn_func,\n)\nApply sequence parallelism slicing to a batch.\nSpecial handling is implemented for integer logits_to_keep, which indicates\nto only keep the last N tokens in the sequence during generation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbatch\ndict[str, torch.Tensor]\nBatch dictionary (e.g., input_ids, attention_mask, etc.).\nrequired\n\n\nlocal_rank\nint\nLocal rank in the sequence parallel group.\nrequired\n\n\nlocal_world_size\nint\nWorld size of the sequence parallel group.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused, but related to above TODO.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[dict[str, torch.Tensor], int, int]\ntuple of: - Batch dictionary with sliced tensors. - The original sequence length before padding. - The number of padding tokens added."
  },
  {
    "objectID": "docs/api/integrations.grokfast.optimizer.html",
    "href": "docs/api/integrations.grokfast.optimizer.html",
    "title": "integrations.grokfast.optimizer",
    "section": "",
    "text": "integrations.grokfast.optimizer\nintegrations.grokfast.optimizer"
  },
  {
    "objectID": "docs/api/utils.callbacks.comet_.html",
    "href": "docs/api/utils.callbacks.comet_.html",
    "title": "utils.callbacks.comet_",
    "section": "",
    "text": "utils.callbacks.comet_\nComet module for trainer callbacks\n\n\n\n\n\nName\nDescription\n\n\n\n\nSaveAxolotlConfigtoCometCallback\nCallback to save axolotl config to comet\n\n\n\n\n\nutils.callbacks.comet_.SaveAxolotlConfigtoCometCallback(axolotl_config_path)\nCallback to save axolotl config to comet"
  },
  {
    "objectID": "docs/api/utils.callbacks.comet_.html#classes",
    "href": "docs/api/utils.callbacks.comet_.html#classes",
    "title": "utils.callbacks.comet_",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nSaveAxolotlConfigtoCometCallback\nCallback to save axolotl config to comet\n\n\n\n\n\nutils.callbacks.comet_.SaveAxolotlConfigtoCometCallback(axolotl_config_path)\nCallback to save axolotl config to comet"
  },
  {
    "objectID": "docs/api/monkeypatch.multipack.html",
    "href": "docs/api/monkeypatch.multipack.html",
    "title": "monkeypatch.multipack",
    "section": "",
    "text": "monkeypatch.multipack\nmonkeypatch.multipack\nmultipack patching for v2 of sample packing"
  },
  {
    "objectID": "docs/api/utils.collators.mm_chat.html",
    "href": "docs/api/utils.collators.mm_chat.html",
    "title": "utils.collators.mm_chat",
    "section": "",
    "text": "utils.collators.mm_chat\nCollators for multi-modal chat messages and packing\n\n\n\n\n\nName\nDescription\n\n\n\n\nMultiModalChatDataCollator\nCollator for multi-modal chat messages\n\n\n\n\n\nutils.collators.mm_chat.MultiModalChatDataCollator(\n    tokenizer,\n    processing_strategy,\n    packing=False,\n    return_tensors='pt',\n    padding=True,\n    pad_to_multiple_of=None,\n)\nCollator for multi-modal chat messages"
  },
  {
    "objectID": "docs/api/utils.collators.mm_chat.html#classes",
    "href": "docs/api/utils.collators.mm_chat.html#classes",
    "title": "utils.collators.mm_chat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMultiModalChatDataCollator\nCollator for multi-modal chat messages\n\n\n\n\n\nutils.collators.mm_chat.MultiModalChatDataCollator(\n    tokenizer,\n    processing_strategy,\n    packing=False,\n    return_tensors='pt',\n    padding=True,\n    pad_to_multiple_of=None,\n)\nCollator for multi-modal chat messages"
  },
  {
    "objectID": "docs/api/utils.tokenization.html",
    "href": "docs/api/utils.tokenization.html",
    "title": "utils.tokenization",
    "section": "",
    "text": "utils.tokenization\nModule for tokenization utilities\n\n\n\n\n\nName\nDescription\n\n\n\n\ncolor_token_for_rl_debug\nHelper function to color tokens based on their type.\n\n\nprocess_tokens_for_rl_debug\nHelper function to process and color tokens.\n\n\n\n\n\nutils.tokenization.color_token_for_rl_debug(\n    decoded_token,\n    encoded_token,\n    color,\n    text_only,\n)\nHelper function to color tokens based on their type.\n\n\n\nutils.tokenization.process_tokens_for_rl_debug(\n    tokens,\n    color,\n    tokenizer,\n    text_only,\n)\nHelper function to process and color tokens."
  },
  {
    "objectID": "docs/api/utils.tokenization.html#functions",
    "href": "docs/api/utils.tokenization.html#functions",
    "title": "utils.tokenization",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ncolor_token_for_rl_debug\nHelper function to color tokens based on their type.\n\n\nprocess_tokens_for_rl_debug\nHelper function to process and color tokens.\n\n\n\n\n\nutils.tokenization.color_token_for_rl_debug(\n    decoded_token,\n    encoded_token,\n    color,\n    text_only,\n)\nHelper function to color tokens based on their type.\n\n\n\nutils.tokenization.process_tokens_for_rl_debug(\n    tokens,\n    color,\n    tokenizer,\n    text_only,\n)\nHelper function to process and color tokens."
  },
  {
    "objectID": "docs/api/prompt_strategies.metharme.html",
    "href": "docs/api/prompt_strategies.metharme.html",
    "title": "prompt_strategies.metharme",
    "section": "",
    "text": "prompt_strategies.metharme\nModule containing the MetharmenPromptTokenizingStrategy and MetharmePrompter class\n\n\n\n\n\nName\nDescription\n\n\n\n\nMetharmePromptTokenizingStrategy\nTokenizing strategy for the Metharme models\n\n\nMetharmePrompter\nPrompter for the Metharme models.\n\n\n\n\n\nprompt_strategies.metharme.MetharmePromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for the Metharme models\n\n\n\nprompt_strategies.metharme.MetharmePrompter(*args, **kwargs)\nPrompter for the Metharme models."
  },
  {
    "objectID": "docs/api/prompt_strategies.metharme.html#classes",
    "href": "docs/api/prompt_strategies.metharme.html#classes",
    "title": "prompt_strategies.metharme",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nMetharmePromptTokenizingStrategy\nTokenizing strategy for the Metharme models\n\n\nMetharmePrompter\nPrompter for the Metharme models.\n\n\n\n\n\nprompt_strategies.metharme.MetharmePromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for the Metharme models\n\n\n\nprompt_strategies.metharme.MetharmePrompter(*args, **kwargs)\nPrompter for the Metharme models."
  },
  {
    "objectID": "docs/api/utils.collators.batching.html",
    "href": "docs/api/utils.collators.batching.html",
    "title": "utils.collators.batching",
    "section": "",
    "text": "utils.collators.batching\nData collators for axolotl to pad labels and position_ids for packed sequences\n\n\n\n\n\nName\nDescription\n\n\n\n\nBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nDataCollatorForSeq2Seq\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\nPretrainingBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nV2BatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\n\n\n\nutils.collators.batching.BatchSamplerDataCollatorForSeq2Seq(\n    tokenizer,\n    model=None,\n    padding=True,\n    max_length=None,\n    pad_to_multiple_of=None,\n    label_pad_token_id=-100,\n    position_pad_token_id=0,\n    return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.DataCollatorForSeq2Seq(\n    tokenizer,\n    model=None,\n    padding=True,\n    max_length=None,\n    pad_to_multiple_of=None,\n    label_pad_token_id=-100,\n    position_pad_token_id=0,\n    return_tensors='pt',\n)\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer\n[PreTrainedTokenizer] or [PreTrainedTokenizerFast]\nThe tokenizer used for encoding the data.\nrequired\n\n\nmodel\n[PreTrainedModel]\nThe model that is being trained. If set and has the prepare_decoder_input_ids_from_labels, use it to prepare the decoder_input_ids This is useful when using label_smoothing to avoid calculating loss twice.\nNone\n\n\npadding\nbool, str or [~utils.PaddingStrategy], optional, defaults to True\nSelect a strategy to pad the returned sequences (according to the model’s padding side and padding index) among: - True or 'longest' (default): Pad to the longest sequence in the batch (or no padding if only a single sequence is provided). - 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. - False or 'do_not_pad': No padding (i.e., can output a batch with sequences of different lengths).\nTrue\n\n\nmax_length\nint, optional\nMaximum length of the returned list and optionally padding length (see above).\nNone\n\n\npad_to_multiple_of\nint, optional\nIf set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).\nNone\n\n\nlabel_pad_token_id\nint, optional, defaults to -100\nThe id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).\n-100\n\n\nreturn_tensors\nstr\nThe type of Tensor to return. Allowable values are “np”, “pt” and “tf”.\n'pt'\n\n\n\n\n\n\n\nutils.collators.batching.PretrainingBatchSamplerDataCollatorForSeq2Seq(\n    *args,\n    multipack_attn=True,\n    **kwargs,\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.V2BatchSamplerDataCollatorForSeq2Seq(\n    tokenizer,\n    model=None,\n    padding=True,\n    max_length=None,\n    pad_to_multiple_of=None,\n    label_pad_token_id=-100,\n    position_pad_token_id=0,\n    return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler"
  },
  {
    "objectID": "docs/api/utils.collators.batching.html#classes",
    "href": "docs/api/utils.collators.batching.html#classes",
    "title": "utils.collators.batching",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nDataCollatorForSeq2Seq\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\nPretrainingBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nV2BatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\n\n\n\nutils.collators.batching.BatchSamplerDataCollatorForSeq2Seq(\n    tokenizer,\n    model=None,\n    padding=True,\n    max_length=None,\n    pad_to_multiple_of=None,\n    label_pad_token_id=-100,\n    position_pad_token_id=0,\n    return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.DataCollatorForSeq2Seq(\n    tokenizer,\n    model=None,\n    padding=True,\n    max_length=None,\n    pad_to_multiple_of=None,\n    label_pad_token_id=-100,\n    position_pad_token_id=0,\n    return_tensors='pt',\n)\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer\n[PreTrainedTokenizer] or [PreTrainedTokenizerFast]\nThe tokenizer used for encoding the data.\nrequired\n\n\nmodel\n[PreTrainedModel]\nThe model that is being trained. If set and has the prepare_decoder_input_ids_from_labels, use it to prepare the decoder_input_ids This is useful when using label_smoothing to avoid calculating loss twice.\nNone\n\n\npadding\nbool, str or [~utils.PaddingStrategy], optional, defaults to True\nSelect a strategy to pad the returned sequences (according to the model’s padding side and padding index) among: - True or 'longest' (default): Pad to the longest sequence in the batch (or no padding if only a single sequence is provided). - 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. - False or 'do_not_pad': No padding (i.e., can output a batch with sequences of different lengths).\nTrue\n\n\nmax_length\nint, optional\nMaximum length of the returned list and optionally padding length (see above).\nNone\n\n\npad_to_multiple_of\nint, optional\nIf set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).\nNone\n\n\nlabel_pad_token_id\nint, optional, defaults to -100\nThe id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).\n-100\n\n\nreturn_tensors\nstr\nThe type of Tensor to return. Allowable values are “np”, “pt” and “tf”.\n'pt'\n\n\n\n\n\n\n\nutils.collators.batching.PretrainingBatchSamplerDataCollatorForSeq2Seq(\n    *args,\n    multipack_attn=True,\n    **kwargs,\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.V2BatchSamplerDataCollatorForSeq2Seq(\n    tokenizer,\n    model=None,\n    padding=True,\n    max_length=None,\n    pad_to_multiple_of=None,\n    label_pad_token_id=-100,\n    position_pad_token_id=0,\n    return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler"
  },
  {
    "objectID": "docs/api/core.trainers.base.html",
    "href": "docs/api/core.trainers.base.html",
    "title": "core.trainers.base",
    "section": "",
    "text": "core.trainers.base\nModule for customized trainers\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlTrainer\nExtend the base Trainer for axolotl helpers\n\n\n\n\n\ncore.trainers.base.AxolotlTrainer(\n    *_args,\n    bench_data_collator=None,\n    eval_data_collator=None,\n    dataset_tags=None,\n    **kwargs,\n)\nExtend the base Trainer for axolotl helpers\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_eval_dataloader\nGet dataloader for evaluation\n\n\nget_train_dataloader\nGet dataloader for training\n\n\nlog\nLog logs on the various objects watching training, including stored metrics.\n\n\npush_to_hub\nOverwrite the push_to_hub method in order to force-add the tags when pushing the\n\n\n\n\n\ncore.trainers.base.AxolotlTrainer.get_eval_dataloader(eval_dataset=None)\nGet dataloader for evaluation\n\n\n\ncore.trainers.base.AxolotlTrainer.get_train_dataloader()\nGet dataloader for training\n\n\n\ncore.trainers.base.AxolotlTrainer.log(logs, start_time=None)\nLog logs on the various objects watching training, including stored metrics.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nlogs\ndict[str, float]\nThe values to log.\nrequired\n\n\nstart_time\nfloat | None\nThe start of training.\nNone\n\n\n\n\n\n\n\ncore.trainers.base.AxolotlTrainer.push_to_hub(*args, **kwargs)\nOverwrite the push_to_hub method in order to force-add the tags when pushing the\nmodel on the Hub. Please refer to ~transformers.Trainer.push_to_hub for more details."
  },
  {
    "objectID": "docs/api/core.trainers.base.html#classes",
    "href": "docs/api/core.trainers.base.html#classes",
    "title": "core.trainers.base",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlTrainer\nExtend the base Trainer for axolotl helpers\n\n\n\n\n\ncore.trainers.base.AxolotlTrainer(\n    *_args,\n    bench_data_collator=None,\n    eval_data_collator=None,\n    dataset_tags=None,\n    **kwargs,\n)\nExtend the base Trainer for axolotl helpers\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_eval_dataloader\nGet dataloader for evaluation\n\n\nget_train_dataloader\nGet dataloader for training\n\n\nlog\nLog logs on the various objects watching training, including stored metrics.\n\n\npush_to_hub\nOverwrite the push_to_hub method in order to force-add the tags when pushing the\n\n\n\n\n\ncore.trainers.base.AxolotlTrainer.get_eval_dataloader(eval_dataset=None)\nGet dataloader for evaluation\n\n\n\ncore.trainers.base.AxolotlTrainer.get_train_dataloader()\nGet dataloader for training\n\n\n\ncore.trainers.base.AxolotlTrainer.log(logs, start_time=None)\nLog logs on the various objects watching training, including stored metrics.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nlogs\ndict[str, float]\nThe values to log.\nrequired\n\n\nstart_time\nfloat | None\nThe start of training.\nNone\n\n\n\n\n\n\n\ncore.trainers.base.AxolotlTrainer.push_to_hub(*args, **kwargs)\nOverwrite the push_to_hub method in order to force-add the tags when pushing the\nmodel on the Hub. Please refer to ~transformers.Trainer.push_to_hub for more details."
  },
  {
    "objectID": "docs/api/core.trainers.grpo.trainer.html",
    "href": "docs/api/core.trainers.grpo.trainer.html",
    "title": "core.trainers.grpo.trainer",
    "section": "",
    "text": "core.trainers.grpo.trainer\nAxolotl GRPO trainers (with and without sequence parallelism handling)\n\n\n\n\n\nName\nDescription\n\n\n\n\nAxolotlGRPOSequenceParallelTrainer\nExtend the base GRPOTrainer for sequence parallelism handling\n\n\nAxolotlGRPOTrainer\nExtend the base GRPOTrainer for axolotl helpers\n\n\n\n\n\ncore.trainers.grpo.trainer.AxolotlGRPOSequenceParallelTrainer(\n    model,\n    reward_funcs,\n    args=None,\n    train_dataset=None,\n    eval_dataset=None,\n    processing_class=None,\n    reward_processing_classes=None,\n    callbacks=None,\n    optimizers=(None, None),\n    peft_config=None,\n    optimizer_cls_and_kwargs=None,\n)\nExtend the base GRPOTrainer for sequence parallelism handling\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_train_dataloader\nGet dataloader for training\n\n\n\n\n\ncore.trainers.grpo.trainer.AxolotlGRPOSequenceParallelTrainer.get_train_dataloader(\n)\nGet dataloader for training\n\n\n\n\n\ncore.trainers.grpo.trainer.AxolotlGRPOTrainer(*args, **kwargs)\nExtend the base GRPOTrainer for axolotl helpers"
  },
  {
    "objectID": "docs/api/core.trainers.grpo.trainer.html#classes",
    "href": "docs/api/core.trainers.grpo.trainer.html#classes",
    "title": "core.trainers.grpo.trainer",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAxolotlGRPOSequenceParallelTrainer\nExtend the base GRPOTrainer for sequence parallelism handling\n\n\nAxolotlGRPOTrainer\nExtend the base GRPOTrainer for axolotl helpers\n\n\n\n\n\ncore.trainers.grpo.trainer.AxolotlGRPOSequenceParallelTrainer(\n    model,\n    reward_funcs,\n    args=None,\n    train_dataset=None,\n    eval_dataset=None,\n    processing_class=None,\n    reward_processing_classes=None,\n    callbacks=None,\n    optimizers=(None, None),\n    peft_config=None,\n    optimizer_cls_and_kwargs=None,\n)\nExtend the base GRPOTrainer for sequence parallelism handling\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_train_dataloader\nGet dataloader for training\n\n\n\n\n\ncore.trainers.grpo.trainer.AxolotlGRPOSequenceParallelTrainer.get_train_dataloader(\n)\nGet dataloader for training\n\n\n\n\n\ncore.trainers.grpo.trainer.AxolotlGRPOTrainer(*args, **kwargs)\nExtend the base GRPOTrainer for axolotl helpers"
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.chatml.html",
    "href": "docs/api/prompt_strategies.dpo.chatml.html",
    "title": "prompt_strategies.dpo.chatml",
    "section": "",
    "text": "prompt_strategies.dpo.chatml\nDPO strategies for chatml\n\n\n\n\n\nName\nDescription\n\n\n\n\nargilla_chat\nfor argilla/dpo-mix-7k conversations\n\n\nicr\nchatml transforms for datasets with system, input, chosen, rejected\n\n\nintel\nFor Intel Orca DPO Pairs\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.dpo.chatml.argilla_chat(cfg, **kwargs)\nfor argilla/dpo-mix-7k conversations\n\n\n\nprompt_strategies.dpo.chatml.icr(cfg, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected\nex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs\n\n\n\nprompt_strategies.dpo.chatml.intel(cfg, **kwargs)\nFor Intel Orca DPO Pairs\n\n\n\nprompt_strategies.dpo.chatml.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations"
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.chatml.html#functions",
    "href": "docs/api/prompt_strategies.dpo.chatml.html#functions",
    "title": "prompt_strategies.dpo.chatml",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nargilla_chat\nfor argilla/dpo-mix-7k conversations\n\n\nicr\nchatml transforms for datasets with system, input, chosen, rejected\n\n\nintel\nFor Intel Orca DPO Pairs\n\n\nultra\nfor ultrafeedback binarized conversations\n\n\n\n\n\nprompt_strategies.dpo.chatml.argilla_chat(cfg, **kwargs)\nfor argilla/dpo-mix-7k conversations\n\n\n\nprompt_strategies.dpo.chatml.icr(cfg, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected\nex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs\n\n\n\nprompt_strategies.dpo.chatml.intel(cfg, **kwargs)\nFor Intel Orca DPO Pairs\n\n\n\nprompt_strategies.dpo.chatml.ultra(cfg, **kwargs)\nfor ultrafeedback binarized conversations"
  },
  {
    "objectID": "docs/api/utils.quantization.html",
    "href": "docs/api/utils.quantization.html",
    "title": "utils.quantization",
    "section": "",
    "text": "utils.quantization\nUtilities for quantization including QAT and PTQ using torchao.\n\n\n\n\n\nName\nDescription\n\n\n\n\nconvert_qat_model_for_ptq\nThis function is used to convert a swap fake-quantized modules in a model\n\n\nget_ptq_config\nThis function is used to build a post-training quantization config.\n\n\nprepare_model_for_qat\nThis function is used to prepare a model for QAT by swapping the model’s linear\n\n\nquantize_model_for_ptq\nThis function is used to quantize a model for post-training quantization.\n\n\n\n\n\nutils.quantization.convert_qat_model_for_ptq(model, *, quantize_embedding=None)\nThis function is used to convert a swap fake-quantized modules in a model\nwhich has been trained with QAT back to the original modules, ready for PTQ.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\n\nThe model to convert.\nrequired\n\n\nquantize_embedding\nbool | None\nWhether to quantize the model’s embedding weights.\nNone\n\n\n\n\n\n\n\nutils.quantization.get_ptq_config(\n    weight_dtype,\n    activation_dtype=None,\n    group_size=None,\n)\nThis function is used to build a post-training quantization config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nweight_dtype\nTorchIntDType\nThe dtype to use for weight quantization.\nrequired\n\n\nactivation_dtype\nTorchIntDType | None\nThe dtype to use for activation quantization.\nNone\n\n\ngroup_size\nint | None\nThe group size to use for weight quantization.\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAOBaseConfig\nThe post-training quantization config.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the activation dtype is not specified and the weight dtype is not int8 or int4, or if the group size is not specified for int8 or int4 weight only quantization.\n\n\n\n\n\n\n\nutils.quantization.prepare_model_for_qat(\n    model,\n    weight_dtype,\n    group_size,\n    activation_dtype=None,\n    quantize_embedding=False,\n)\nThis function is used to prepare a model for QAT by swapping the model’s linear\nlayers with fake quantized linear layers, and optionally the embedding weights with\nfake quantized embedding weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\n\nThe model to quantize.\nrequired\n\n\nweight_dtype\nTorchIntDType\nThe dtype to use for weight quantization.\nrequired\n\n\ngroup_size\nint\nThe group size to use for weight quantization.\nrequired\n\n\nactivation_dtype\nTorchIntDType | None\nThe dtype to use for activation quantization.\nNone\n\n\nquantize_embedding\nbool\nWhether to quantize the model’s embedding weights.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the activation/weight dtype combination is invalid.\n\n\n\n\n\n\n\nutils.quantization.quantize_model_for_ptq(\n    model,\n    weight_dtype,\n    group_size=None,\n    activation_dtype=None,\n    quantize_embedding=None,\n)\nThis function is used to quantize a model for post-training quantization.\nIt swaps the model’s linear layers with fake quantized linear layers.\nIf quantize_embedding is True, it will also swap the model’s embedding weights with fake quantized embedding weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\n\nThe model to quantize.\nrequired\n\n\nweight_dtype\nTorchIntDType\nThe dtype to use for weight quantization.\nrequired\n\n\ngroup_size\nint | None\nThe group size to use for weight quantization.\nNone\n\n\nactivation_dtype\nTorchIntDType | None\nThe dtype to use for activation quantization.\nNone\n\n\nquantize_embedding\nbool | None\nWhether to quantize the model’s embedding weights.\nNone"
  },
  {
    "objectID": "docs/api/utils.quantization.html#functions",
    "href": "docs/api/utils.quantization.html#functions",
    "title": "utils.quantization",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nconvert_qat_model_for_ptq\nThis function is used to convert a swap fake-quantized modules in a model\n\n\nget_ptq_config\nThis function is used to build a post-training quantization config.\n\n\nprepare_model_for_qat\nThis function is used to prepare a model for QAT by swapping the model’s linear\n\n\nquantize_model_for_ptq\nThis function is used to quantize a model for post-training quantization.\n\n\n\n\n\nutils.quantization.convert_qat_model_for_ptq(model, *, quantize_embedding=None)\nThis function is used to convert a swap fake-quantized modules in a model\nwhich has been trained with QAT back to the original modules, ready for PTQ.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\n\nThe model to convert.\nrequired\n\n\nquantize_embedding\nbool | None\nWhether to quantize the model’s embedding weights.\nNone\n\n\n\n\n\n\n\nutils.quantization.get_ptq_config(\n    weight_dtype,\n    activation_dtype=None,\n    group_size=None,\n)\nThis function is used to build a post-training quantization config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nweight_dtype\nTorchIntDType\nThe dtype to use for weight quantization.\nrequired\n\n\nactivation_dtype\nTorchIntDType | None\nThe dtype to use for activation quantization.\nNone\n\n\ngroup_size\nint | None\nThe group size to use for weight quantization.\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAOBaseConfig\nThe post-training quantization config.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the activation dtype is not specified and the weight dtype is not int8 or int4, or if the group size is not specified for int8 or int4 weight only quantization.\n\n\n\n\n\n\n\nutils.quantization.prepare_model_for_qat(\n    model,\n    weight_dtype,\n    group_size,\n    activation_dtype=None,\n    quantize_embedding=False,\n)\nThis function is used to prepare a model for QAT by swapping the model’s linear\nlayers with fake quantized linear layers, and optionally the embedding weights with\nfake quantized embedding weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\n\nThe model to quantize.\nrequired\n\n\nweight_dtype\nTorchIntDType\nThe dtype to use for weight quantization.\nrequired\n\n\ngroup_size\nint\nThe group size to use for weight quantization.\nrequired\n\n\nactivation_dtype\nTorchIntDType | None\nThe dtype to use for activation quantization.\nNone\n\n\nquantize_embedding\nbool\nWhether to quantize the model’s embedding weights.\nFalse\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the activation/weight dtype combination is invalid.\n\n\n\n\n\n\n\nutils.quantization.quantize_model_for_ptq(\n    model,\n    weight_dtype,\n    group_size=None,\n    activation_dtype=None,\n    quantize_embedding=None,\n)\nThis function is used to quantize a model for post-training quantization.\nIt swaps the model’s linear layers with fake quantized linear layers.\nIf quantize_embedding is True, it will also swap the model’s embedding weights with fake quantized embedding weights.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\n\nThe model to quantize.\nrequired\n\n\nweight_dtype\nTorchIntDType\nThe dtype to use for weight quantization.\nrequired\n\n\ngroup_size\nint | None\nThe group size to use for weight quantization.\nNone\n\n\nactivation_dtype\nTorchIntDType | None\nThe dtype to use for activation quantization.\nNone\n\n\nquantize_embedding\nbool | None\nWhether to quantize the model’s embedding weights.\nNone"
  },
  {
    "objectID": "docs/api/prompt_strategies.llama2_chat.html",
    "href": "docs/api/prompt_strategies.llama2_chat.html",
    "title": "prompt_strategies.llama2_chat",
    "section": "",
    "text": "prompt_strategies.llama2_chat\nPrompt Strategy for finetuning Llama2 chat models\nsee also https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/generation.py#L213 for ma reference implementation.\nThis implementation is based on the Vicuna PR and the fastchat repo, see also:\nhttps://github.com/lm-sys/FastChat/blob/cdd7730686cb1bf9ae2b768ee171bdf7d1ff04f3/fastchat/conversation.py#L847\nUse dataset type: “llama2_chat” in conig.yml to use this prompt style.\nE.g. in the config.yml:\ndatasets:\n  - path: llama_finetune_train.jsonl\n    type: llama2_chat\nThe dataset itself should look like this:\n{'conversations':[{\"from\": \"human\", \"value\": \"Who are you?\"}, {\"from\": \"gpt\", \"value\": \"I am Vicuna\"},...]}\nin a jsonl file. The first message should be from the human, the second from gpt.\nFor a custom system message, the first “from” can be “system” (followed by alternating “human” and “gpt” turns).\nImportant: Don’t use “special_tokens:” in your config.yml if you are not sure what you are doing!\n\n\n\n\n\nName\nDescription\n\n\n\n\nLLama2ChatTokenizingStrategy\nTokenizing strategy for Llama2 prompts.\n\n\nLlama2ChatConversation\nA class that manages prompt templates and keeps all conversation history.\n\n\nLlama2ChatPrompter\nA prompter that generates prompts for Llama2 models.\n\n\n\n\n\nprompt_strategies.llama2_chat.LLama2ChatTokenizingStrategy(*args, **kwargs)\nTokenizing strategy for Llama2 prompts.\nadapted from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatConversation(\n    name='llama2',\n    system=\"[INST] <<SYS>>\\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\\n\\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.\\n<</SYS>>\\n\\n\",\n    roles=('[INST]', '[/INST]'),\n    messages=list(),\n    offset=0,\n)\nA class that manages prompt templates and keeps all conversation history.\ncopied from https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n\n\n\n\n\nName\nDescription\n\n\n\n\nappend_message\nAppend a new message.\n\n\nget_prompt\nGet the prompt for generation.\n\n\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatConversation.append_message(\n    role,\n    message,\n)\nAppend a new message.\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatConversation.get_prompt()\nGet the prompt for generation.\n\n\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatPrompter()\nA prompter that generates prompts for Llama2 models."
  },
  {
    "objectID": "docs/api/prompt_strategies.llama2_chat.html#classes",
    "href": "docs/api/prompt_strategies.llama2_chat.html#classes",
    "title": "prompt_strategies.llama2_chat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nLLama2ChatTokenizingStrategy\nTokenizing strategy for Llama2 prompts.\n\n\nLlama2ChatConversation\nA class that manages prompt templates and keeps all conversation history.\n\n\nLlama2ChatPrompter\nA prompter that generates prompts for Llama2 models.\n\n\n\n\n\nprompt_strategies.llama2_chat.LLama2ChatTokenizingStrategy(*args, **kwargs)\nTokenizing strategy for Llama2 prompts.\nadapted from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatConversation(\n    name='llama2',\n    system=\"[INST] <<SYS>>\\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\\n\\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.\\n<</SYS>>\\n\\n\",\n    roles=('[INST]', '[/INST]'),\n    messages=list(),\n    offset=0,\n)\nA class that manages prompt templates and keeps all conversation history.\ncopied from https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n\n\n\n\n\nName\nDescription\n\n\n\n\nappend_message\nAppend a new message.\n\n\nget_prompt\nGet the prompt for generation.\n\n\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatConversation.append_message(\n    role,\n    message,\n)\nAppend a new message.\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatConversation.get_prompt()\nGet the prompt for generation.\n\n\n\n\n\nprompt_strategies.llama2_chat.Llama2ChatPrompter()\nA prompter that generates prompts for Llama2 models."
  },
  {
    "objectID": "docs/api/cli.quantize.html",
    "href": "docs/api/cli.quantize.html",
    "title": "cli.quantize",
    "section": "",
    "text": "cli.quantize\nCLI to post-training quantize a model using torchao\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_quantize\nQuantizes a model’s model’s weights\n\n\n\n\n\ncli.quantize.do_quantize(config, cli_args)\nQuantizes a model’s model’s weights\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nThe path to the config file\nrequired\n\n\ncli_args\ndict\nAdditional command-line arguments\nrequired"
  },
  {
    "objectID": "docs/api/cli.quantize.html#functions",
    "href": "docs/api/cli.quantize.html#functions",
    "title": "cli.quantize",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_quantize\nQuantizes a model’s model’s weights\n\n\n\n\n\ncli.quantize.do_quantize(config, cli_args)\nQuantizes a model’s model’s weights\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nThe path to the config file\nrequired\n\n\ncli_args\ndict\nAdditional command-line arguments\nrequired"
  },
  {
    "objectID": "docs/api/prompt_strategies.alpaca_instruct.html",
    "href": "docs/api/prompt_strategies.alpaca_instruct.html",
    "title": "prompt_strategies.alpaca_instruct",
    "section": "",
    "text": "prompt_strategies.alpaca_instruct\nprompt_strategies.alpaca_instruct\nModule loading the AlpacaInstructPromptTokenizingStrategy class"
  },
  {
    "objectID": "docs/api/cli.vllm_serve.html",
    "href": "docs/api/cli.vllm_serve.html",
    "title": "cli.vllm_serve",
    "section": "",
    "text": "cli.vllm_serve\nCLI to start the vllm server for online RL\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_vllm_serve\nStarts the VLLM server for serving LLM models used for online RL\n\n\n\n\n\ncli.vllm_serve.do_vllm_serve(config, cli_args)\nStarts the VLLM server for serving LLM models used for online RL\nArgs\n:param cfg: Parsed doct of the YAML config\n:param cli_args: dict of additional command-line arguments of type VllmServeCliArgs\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nprocess_id\n\nthe process id of the started VLLM server"
  },
  {
    "objectID": "docs/api/cli.vllm_serve.html#functions",
    "href": "docs/api/cli.vllm_serve.html#functions",
    "title": "cli.vllm_serve",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_vllm_serve\nStarts the VLLM server for serving LLM models used for online RL\n\n\n\n\n\ncli.vllm_serve.do_vllm_serve(config, cli_args)\nStarts the VLLM server for serving LLM models used for online RL\nArgs\n:param cfg: Parsed doct of the YAML config\n:param cli_args: dict of additional command-line arguments of type VllmServeCliArgs\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nprocess_id\n\nthe process id of the started VLLM server"
  },
  {
    "objectID": "docs/api/prompt_strategies.alpaca_chat.html",
    "href": "docs/api/prompt_strategies.alpaca_chat.html",
    "title": "prompt_strategies.alpaca_chat",
    "section": "",
    "text": "prompt_strategies.alpaca_chat\nModule for Alpaca prompt strategy classes\n\n\n\n\n\nName\nDescription\n\n\n\n\nAlpacaChatPrompter\nAlpaca Chat Prompter extending the system prompt to for chat-instruct answers\n\n\nAlpacaConcisePrompter\nAlpaca Prompter extending the system prompt to ask for concise chat-instruct answers\n\n\nAlpacaQAPromptTokenizingStrategy\nTokenizing strategy for AlpacaQA\n\n\nCamelAIPromptTokenizingStrategy\nTokenizing strategy for CamelAI datasets\n\n\nNoSystemPrompter\nNull Prompter with no system prompts\n\n\n\n\n\nprompt_strategies.alpaca_chat.AlpacaChatPrompter()\nAlpaca Chat Prompter extending the system prompt to for chat-instruct answers\n\n\n\nprompt_strategies.alpaca_chat.AlpacaConcisePrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAlpaca Prompter extending the system prompt to ask for concise chat-instruct answers\n\n\n\nprompt_strategies.alpaca_chat.AlpacaQAPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for AlpacaQA\n\n\n\nprompt_strategies.alpaca_chat.CamelAIPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for CamelAI datasets\n\n\n\nprompt_strategies.alpaca_chat.NoSystemPrompter()\nNull Prompter with no system prompts"
  },
  {
    "objectID": "docs/api/prompt_strategies.alpaca_chat.html#classes",
    "href": "docs/api/prompt_strategies.alpaca_chat.html#classes",
    "title": "prompt_strategies.alpaca_chat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nAlpacaChatPrompter\nAlpaca Chat Prompter extending the system prompt to for chat-instruct answers\n\n\nAlpacaConcisePrompter\nAlpaca Prompter extending the system prompt to ask for concise chat-instruct answers\n\n\nAlpacaQAPromptTokenizingStrategy\nTokenizing strategy for AlpacaQA\n\n\nCamelAIPromptTokenizingStrategy\nTokenizing strategy for CamelAI datasets\n\n\nNoSystemPrompter\nNull Prompter with no system prompts\n\n\n\n\n\nprompt_strategies.alpaca_chat.AlpacaChatPrompter()\nAlpaca Chat Prompter extending the system prompt to for chat-instruct answers\n\n\n\nprompt_strategies.alpaca_chat.AlpacaConcisePrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAlpaca Prompter extending the system prompt to ask for concise chat-instruct answers\n\n\n\nprompt_strategies.alpaca_chat.AlpacaQAPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for AlpacaQA\n\n\n\nprompt_strategies.alpaca_chat.CamelAIPromptTokenizingStrategy(\n    prompter,\n    tokenizer,\n    train_on_inputs=False,\n    sequence_len=2048,\n)\nTokenizing strategy for CamelAI datasets\n\n\n\nprompt_strategies.alpaca_chat.NoSystemPrompter()\nNull Prompter with no system prompts"
  },
  {
    "objectID": "docs/api/prompt_strategies.bradley_terry.llama3.html",
    "href": "docs/api/prompt_strategies.bradley_terry.llama3.html",
    "title": "prompt_strategies.bradley_terry.llama3",
    "section": "",
    "text": "prompt_strategies.bradley_terry.llama3\nchatml transforms for datasets with system, input, chosen, rejected to match llama3 chat template\n\n\n\n\n\nName\nDescription\n\n\n\n\nicr\nchatml transforms for datasets with system, input, chosen, rejected\n\n\n\n\n\nprompt_strategies.bradley_terry.llama3.icr(cfg, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected\nex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs"
  },
  {
    "objectID": "docs/api/prompt_strategies.bradley_terry.llama3.html#functions",
    "href": "docs/api/prompt_strategies.bradley_terry.llama3.html#functions",
    "title": "prompt_strategies.bradley_terry.llama3",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nicr\nchatml transforms for datasets with system, input, chosen, rejected\n\n\n\n\n\nprompt_strategies.bradley_terry.llama3.icr(cfg, **kwargs)\nchatml transforms for datasets with system, input, chosen, rejected\nex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs"
  },
  {
    "objectID": "docs/api/utils.dict.html",
    "href": "docs/api/utils.dict.html",
    "title": "utils.dict",
    "section": "",
    "text": "utils.dict\nModule containing the DictDefault class\n\n\n\n\n\nName\nDescription\n\n\n\n\nDictDefault\nA Dict that returns None instead of returning empty Dict for missing keys.\n\n\n\n\n\nutils.dict.DictDefault()\nA Dict that returns None instead of returning empty Dict for missing keys."
  },
  {
    "objectID": "docs/api/utils.dict.html#classes",
    "href": "docs/api/utils.dict.html#classes",
    "title": "utils.dict",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nDictDefault\nA Dict that returns None instead of returning empty Dict for missing keys.\n\n\n\n\n\nutils.dict.DictDefault()\nA Dict that returns None instead of returning empty Dict for missing keys."
  },
  {
    "objectID": "docs/api/prompt_strategies.dpo.user_defined.html",
    "href": "docs/api/prompt_strategies.dpo.user_defined.html",
    "title": "prompt_strategies.dpo.user_defined",
    "section": "",
    "text": "prompt_strategies.dpo.user_defined\nprompt_strategies.dpo.user_defined\nUser-defined DPO strategies"
  },
  {
    "objectID": "docs/api/monkeypatch.llama_attn_hijack_xformers.html",
    "href": "docs/api/monkeypatch.llama_attn_hijack_xformers.html",
    "title": "monkeypatch.llama_attn_hijack_xformers",
    "section": "",
    "text": "monkeypatch.llama_attn_hijack_xformers\nmonkeypatch.llama_attn_hijack_xformers\nDirectly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments"
  },
  {
    "objectID": "docs/api/utils.lora.html",
    "href": "docs/api/utils.lora.html",
    "title": "utils.lora",
    "section": "",
    "text": "utils.lora\nmodule to get the state dict of a merged lora model\n\n\n\n\n\nName\nDescription\n\n\n\n\nget_lora_merged_state_dict\nCreate and return a state_dict that has the LoRA deltas\n\n\n\n\n\nutils.lora.get_lora_merged_state_dict(model)\nCreate and return a state_dict that has the LoRA deltas\nmerged into the base model’s weights, without modifying model in place.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\ntorch.nn.Module\nA model that has LoRA/PEFT adapters attached.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\ndict\ndict\nA state_dict of the merged parameters."
  },
  {
    "objectID": "docs/api/utils.lora.html#functions",
    "href": "docs/api/utils.lora.html#functions",
    "title": "utils.lora",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nget_lora_merged_state_dict\nCreate and return a state_dict that has the LoRA deltas\n\n\n\n\n\nutils.lora.get_lora_merged_state_dict(model)\nCreate and return a state_dict that has the LoRA deltas\nmerged into the base model’s weights, without modifying model in place.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\ntorch.nn.Module\nA model that has LoRA/PEFT adapters attached.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\ndict\ndict\nA state_dict of the merged parameters."
  },
  {
    "objectID": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html",
    "href": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html",
    "title": "monkeypatch.gradient_checkpointing.offload_cpu",
    "section": "",
    "text": "monkeypatch.gradient_checkpointing.offload_cpu\nCPU offloaded checkpointing\n\n\n\n\n\nName\nDescription\n\n\n\n\nCPU_Offloaded_Gradient_Checkpointer\nSaves VRAM by smartly offloading to RAM.\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CPU_Offloaded_Gradient_Checkpointer(\n)\nSaves VRAM by smartly offloading to RAM.\nTiny hit to performance, since we mask the movement via non blocking calls."
  },
  {
    "objectID": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html#classes",
    "href": "docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html#classes",
    "title": "monkeypatch.gradient_checkpointing.offload_cpu",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nCPU_Offloaded_Gradient_Checkpointer\nSaves VRAM by smartly offloading to RAM.\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_cpu.CPU_Offloaded_Gradient_Checkpointer(\n)\nSaves VRAM by smartly offloading to RAM.\nTiny hit to performance, since we mask the movement via non blocking calls."
  },
  {
    "objectID": "docs/api/utils.trainer.html",
    "href": "docs/api/utils.trainer.html",
    "title": "utils.trainer",
    "section": "",
    "text": "utils.trainer\nModule containing the Trainer class and related functions\n\n\n\n\n\nName\nDescription\n\n\n\n\nadd_pose_position_ids\nuse the PoSE technique to extend the context length by randomly skipping\n\n\nadd_position_ids\nHandle both single-example and batched data.\n\n\ndrop_long_seq\nDrop samples whose sequence length is either too long (> sequence_len)\n\n\nsetup_trainer\nHelper method for instantiating and building a (causal or RLHF) trainer.\n\n\n\n\n\nutils.trainer.add_pose_position_ids(\n    sample,\n    max_context_len=32768,\n    split_on_token_ids=None,\n    chunks=2,\n)\nuse the PoSE technique to extend the context length by randomly skipping\npositions in the context. We only want to skip right before tokens in\nthe split_on_token_ids list. We should attempt to randomly distribute\nthe skips, but we don’t need the final position_ids to be the full\ncontext_len. There may be multiple turns in the context, so we want to\nmake sure we take into account the maximum possible number of skips\nremaining in each sample.\n\n\n\nutils.trainer.add_position_ids(sample)\nHandle both single-example and batched data.\n- single example: sample[‘input_ids’] is a list[int]\n- batched data: sample[‘input_ids’] is a list[list[int]]\n\n\n\nutils.trainer.drop_long_seq(sample, sequence_len=2048, min_sequence_len=2)\nDrop samples whose sequence length is either too long (> sequence_len)\nor too short (< min_sequence_len).\nWorks for both single-example (list[int]) or batched (list[list[int]]).\n\n\n\nutils.trainer.setup_trainer(\n    cfg,\n    train_dataset,\n    eval_dataset,\n    model,\n    tokenizer,\n    processor,\n    total_num_steps,\n    model_ref=None,\n    peft_config=None,\n)\nHelper method for instantiating and building a (causal or RLHF) trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\n\nAxolotl config object containing training parameters.\nrequired\n\n\ntrain_dataset\n\nDataset to use for training.\nrequired\n\n\neval_dataset\n\nDataset to use for evaluation.\nrequired\n\n\nmodel\n\nThe model to train.\nrequired\n\n\ntokenizer\n\nTokenizer for processing text input.\nrequired\n\n\nprocessor\n\nProcessor for data preparation.\nrequired\n\n\ntotal_num_steps\n\nThe total number of training steps.\nrequired\n\n\nmodel_ref\n\nOptional reference model for RLHF training. Default is None.\nNone\n\n\npeft_config\n\nOptional PEFT (Parameter-Efficient Fine-Tuning) configuration. Default is None.\nNone\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\n\nA trainer instance (either HFRLTrainer or HFCausalTrainer) configured based on the provided parameters."
  },
  {
    "objectID": "docs/api/utils.trainer.html#functions",
    "href": "docs/api/utils.trainer.html#functions",
    "title": "utils.trainer",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nadd_pose_position_ids\nuse the PoSE technique to extend the context length by randomly skipping\n\n\nadd_position_ids\nHandle both single-example and batched data.\n\n\ndrop_long_seq\nDrop samples whose sequence length is either too long (> sequence_len)\n\n\nsetup_trainer\nHelper method for instantiating and building a (causal or RLHF) trainer.\n\n\n\n\n\nutils.trainer.add_pose_position_ids(\n    sample,\n    max_context_len=32768,\n    split_on_token_ids=None,\n    chunks=2,\n)\nuse the PoSE technique to extend the context length by randomly skipping\npositions in the context. We only want to skip right before tokens in\nthe split_on_token_ids list. We should attempt to randomly distribute\nthe skips, but we don’t need the final position_ids to be the full\ncontext_len. There may be multiple turns in the context, so we want to\nmake sure we take into account the maximum possible number of skips\nremaining in each sample.\n\n\n\nutils.trainer.add_position_ids(sample)\nHandle both single-example and batched data.\n- single example: sample[‘input_ids’] is a list[int]\n- batched data: sample[‘input_ids’] is a list[list[int]]\n\n\n\nutils.trainer.drop_long_seq(sample, sequence_len=2048, min_sequence_len=2)\nDrop samples whose sequence length is either too long (> sequence_len)\nor too short (< min_sequence_len).\nWorks for both single-example (list[int]) or batched (list[list[int]]).\n\n\n\nutils.trainer.setup_trainer(\n    cfg,\n    train_dataset,\n    eval_dataset,\n    model,\n    tokenizer,\n    processor,\n    total_num_steps,\n    model_ref=None,\n    peft_config=None,\n)\nHelper method for instantiating and building a (causal or RLHF) trainer.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\n\nAxolotl config object containing training parameters.\nrequired\n\n\ntrain_dataset\n\nDataset to use for training.\nrequired\n\n\neval_dataset\n\nDataset to use for evaluation.\nrequired\n\n\nmodel\n\nThe model to train.\nrequired\n\n\ntokenizer\n\nTokenizer for processing text input.\nrequired\n\n\nprocessor\n\nProcessor for data preparation.\nrequired\n\n\ntotal_num_steps\n\nThe total number of training steps.\nrequired\n\n\nmodel_ref\n\nOptional reference model for RLHF training. Default is None.\nNone\n\n\npeft_config\n\nOptional PEFT (Parameter-Efficient Fine-Tuning) configuration. Default is None.\nNone\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\n\nA trainer instance (either HFRLTrainer or HFCausalTrainer) configured based on the provided parameters."
  },
  {
    "objectID": "docs/api/prompt_strategies.orcamini.html",
    "href": "docs/api/prompt_strategies.orcamini.html",
    "title": "prompt_strategies.orcamini",
    "section": "",
    "text": "prompt_strategies.orcamini\nPrompt Strategy for finetuning Orca Mini (v2) models\nsee also https://huggingface.co/psmathur/orca_mini_v2_7b for more information\nUse dataset type: orcamini in conig.yml to use this prompt style.\nCompared to the alpaca_w_system.open_orca dataset type,\nthis one specifies the system prompt with “### System:”.\nNot suited/tested for multiple-turn conversations without further adjustments.\n\n\n\n\n\nName\nDescription\n\n\n\n\nOrcaMiniPrompter\nAdjusted Prompter for Orca Mini (v2) datasets\n\n\n\n\n\nprompt_strategies.orcamini.OrcaMiniPrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAdjusted Prompter for Orca Mini (v2) datasets"
  },
  {
    "objectID": "docs/api/prompt_strategies.orcamini.html#classes",
    "href": "docs/api/prompt_strategies.orcamini.html#classes",
    "title": "prompt_strategies.orcamini",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nOrcaMiniPrompter\nAdjusted Prompter for Orca Mini (v2) datasets\n\n\n\n\n\nprompt_strategies.orcamini.OrcaMiniPrompter(\n    prompt_style=PromptStyle.INSTRUCT.value,\n)\nAdjusted Prompter for Orca Mini (v2) datasets"
  },
  {
    "objectID": "docs/api/prompt_strategies.messages.chat.html",
    "href": "docs/api/prompt_strategies.messages.chat.html",
    "title": "prompt_strategies.messages.chat",
    "section": "",
    "text": "prompt_strategies.messages.chat\nChat dataset wrapping strategy for new internal messages representations\n\n\n\n\n\nName\nDescription\n\n\n\n\nChatMessageDatasetWrappingStrategy\nChat dataset wrapping strategy for new internal messages representations\n\n\n\n\n\nprompt_strategies.messages.chat.ChatMessageDatasetWrappingStrategy(\n    processor,\n    message_transform=None,\n    formatter=None,\n    **kwargs,\n)\nChat dataset wrapping strategy for new internal messages representations"
  },
  {
    "objectID": "docs/api/prompt_strategies.messages.chat.html#classes",
    "href": "docs/api/prompt_strategies.messages.chat.html#classes",
    "title": "prompt_strategies.messages.chat",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nChatMessageDatasetWrappingStrategy\nChat dataset wrapping strategy for new internal messages representations\n\n\n\n\n\nprompt_strategies.messages.chat.ChatMessageDatasetWrappingStrategy(\n    processor,\n    message_transform=None,\n    formatter=None,\n    **kwargs,\n)\nChat dataset wrapping strategy for new internal messages representations"
  },
  {
    "objectID": "docs/api/monkeypatch.lora_kernels.html",
    "href": "docs/api/monkeypatch.lora_kernels.html",
    "title": "monkeypatch.lora_kernels",
    "section": "",
    "text": "monkeypatch.lora_kernels\nModule for patching custom LoRA Triton kernels and torch.autograd functions.\n\n\n\n\n\nName\nDescription\n\n\n\n\nFakeMLP\nplaceholder MLP for triton patching\n\n\n\n\n\nmonkeypatch.lora_kernels.FakeMLP(gate_proj, up_proj, down_proj)\nplaceholder MLP for triton patching\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_lora_kernel_patches\nApplies optimized Triton kernel patches to a PEFT model.\n\n\nget_attention_cls_from_config\nGet the appropriate attention class by inspecting the model config.\n\n\noriginal_apply_o\nOriginal implementation of output projection without optimizations.\n\n\noriginal_apply_qkv\nOriginal implementation of QKV projection without optimizations.\n\n\npatch_self_attn_lora\nGiven an axolotl config, this method patches the inferred attention class forward\n\n\n\n\n\nmonkeypatch.lora_kernels.apply_lora_kernel_patches(model, cfg)\nApplies optimized Triton kernel patches to a PEFT model.\nPatches a PEFT model with optimized implementations for MLP and attention\ncomputations. The optimizations include custom Triton kernels for activation\nfunctions and specialized autograd functions for LoRA computations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model to be patched with optimized kernels.\nrequired\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nPeftModelForCausalLM\nPeftModelForCausalLM\nThe patched model with optimized kernels.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTypeError\nIf the provided model is not a PeftModelForCausalLM.\n\n\n\nNotImplementedError\nIf the model type is not supported.\n\n\n\nAssertionError\nIf multiple adapters are active (currently unsupported).\n\n\n\n\n\n\nThe optimizations require LoRA adapters with no dropout and no bias terms. The\nfunction will skip patching if these conditions aren’t met.\n\n\n\n\nmonkeypatch.lora_kernels.get_attention_cls_from_config(cfg)\nGet the appropriate attention class by inspecting the model config.\nUses dynamic import to support any model architecture that follows\nthe standard transformers naming convention.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nType[nn.Module]\nThe appropriate attention class for the model.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf base_model not specified or attention class cannot be imported\n\n\n\nImportError\nIf the model module or attention class doesn’t exist\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_o(self, hidden_states)\nOriginal implementation of output projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim]`.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nThe output projection result.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_qkv(self, hidden_states)\nOriginal implementation of QKV projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nA tuple (query_states, key_states, value_states) containing the projected states for query, key, and value.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.patch_self_attn_lora(cfg)\nGiven an axolotl config, this method patches the inferred attention class forward\npass with optimized LoRA implementations.\nIt modifies the attention class to use optimized QKV and output projections. The\noriginal implementation is preserved and can be restored if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf the required code blocks are not found in the attention implementation."
  },
  {
    "objectID": "docs/api/monkeypatch.lora_kernels.html#classes",
    "href": "docs/api/monkeypatch.lora_kernels.html#classes",
    "title": "monkeypatch.lora_kernels",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nFakeMLP\nplaceholder MLP for triton patching\n\n\n\n\n\nmonkeypatch.lora_kernels.FakeMLP(gate_proj, up_proj, down_proj)\nplaceholder MLP for triton patching"
  },
  {
    "objectID": "docs/api/monkeypatch.lora_kernels.html#functions",
    "href": "docs/api/monkeypatch.lora_kernels.html#functions",
    "title": "monkeypatch.lora_kernels",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\napply_lora_kernel_patches\nApplies optimized Triton kernel patches to a PEFT model.\n\n\nget_attention_cls_from_config\nGet the appropriate attention class by inspecting the model config.\n\n\noriginal_apply_o\nOriginal implementation of output projection without optimizations.\n\n\noriginal_apply_qkv\nOriginal implementation of QKV projection without optimizations.\n\n\npatch_self_attn_lora\nGiven an axolotl config, this method patches the inferred attention class forward\n\n\n\n\n\nmonkeypatch.lora_kernels.apply_lora_kernel_patches(model, cfg)\nApplies optimized Triton kernel patches to a PEFT model.\nPatches a PEFT model with optimized implementations for MLP and attention\ncomputations. The optimizations include custom Triton kernels for activation\nfunctions and specialized autograd functions for LoRA computations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model to be patched with optimized kernels.\nrequired\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nPeftModelForCausalLM\nPeftModelForCausalLM\nThe patched model with optimized kernels.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTypeError\nIf the provided model is not a PeftModelForCausalLM.\n\n\n\nNotImplementedError\nIf the model type is not supported.\n\n\n\nAssertionError\nIf multiple adapters are active (currently unsupported).\n\n\n\n\n\n\nThe optimizations require LoRA adapters with no dropout and no bias terms. The\nfunction will skip patching if these conditions aren’t met.\n\n\n\n\nmonkeypatch.lora_kernels.get_attention_cls_from_config(cfg)\nGet the appropriate attention class by inspecting the model config.\nUses dynamic import to support any model architecture that follows\nthe standard transformers naming convention.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nType[nn.Module]\nThe appropriate attention class for the model.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf base_model not specified or attention class cannot be imported\n\n\n\nImportError\nIf the model module or attention class doesn’t exist\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_o(self, hidden_states)\nOriginal implementation of output projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim]`.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nThe output projection result.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_qkv(self, hidden_states)\nOriginal implementation of QKV projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nA tuple (query_states, key_states, value_states) containing the projected states for query, key, and value.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.patch_self_attn_lora(cfg)\nGiven an axolotl config, this method patches the inferred attention class forward\npass with optimized LoRA implementations.\nIt modifies the attention class to use optimized QKV and output projections. The\noriginal implementation is preserved and can be restored if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf the required code blocks are not found in the attention implementation."
  },
  {
    "objectID": "docs/api/utils.schemas.enums.html",
    "href": "docs/api/utils.schemas.enums.html",
    "title": "utils.schemas.enums",
    "section": "",
    "text": "utils.schemas.enums\nEnums for Axolotl input config\n\n\n\n\n\nName\nDescription\n\n\n\n\nChatTemplate\nChat templates configuration subset\n\n\nCustomSupportedOptimizers\nCustom supported optimizers\n\n\nRLType\nRL trainer type configuration subset\n\n\nRingAttnFunc\nEnum class for supported ring-flash-attn implementations\n\n\nTorchIntDType\nTorch integer data types - getattr guards against torch < 2.6 which does not support int4\n\n\n\n\n\nutils.schemas.enums.ChatTemplate()\nChat templates configuration subset\n\n\n\nutils.schemas.enums.CustomSupportedOptimizers()\nCustom supported optimizers\n\n\n\nutils.schemas.enums.RLType()\nRL trainer type configuration subset\n\n\n\nutils.schemas.enums.RingAttnFunc()\nEnum class for supported ring-flash-attn implementations\n\n\n\nutils.schemas.enums.TorchIntDType()\nTorch integer data types - getattr guards against torch < 2.6 which does not support int4"
  },
  {
    "objectID": "docs/api/utils.schemas.enums.html#classes",
    "href": "docs/api/utils.schemas.enums.html#classes",
    "title": "utils.schemas.enums",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nChatTemplate\nChat templates configuration subset\n\n\nCustomSupportedOptimizers\nCustom supported optimizers\n\n\nRLType\nRL trainer type configuration subset\n\n\nRingAttnFunc\nEnum class for supported ring-flash-attn implementations\n\n\nTorchIntDType\nTorch integer data types - getattr guards against torch < 2.6 which does not support int4\n\n\n\n\n\nutils.schemas.enums.ChatTemplate()\nChat templates configuration subset\n\n\n\nutils.schemas.enums.CustomSupportedOptimizers()\nCustom supported optimizers\n\n\n\nutils.schemas.enums.RLType()\nRL trainer type configuration subset\n\n\n\nutils.schemas.enums.RingAttnFunc()\nEnum class for supported ring-flash-attn implementations\n\n\n\nutils.schemas.enums.TorchIntDType()\nTorch integer data types - getattr guards against torch < 2.6 which does not support int4"
  },
  {
    "objectID": "docs/api/cli.config.html",
    "href": "docs/api/cli.config.html",
    "title": "cli.config",
    "section": "",
    "text": "cli.config\nConfiguration loading and processing.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncheck_remote_config\nFirst, determines if the passed config is a valid HTTPS URL. Then, attempts to query\n\n\nchoose_config\nHelper method for choosing a axolotl config YAML file (considering only files\n\n\nload_cfg\nLoads the axolotl configuration stored at config, validates it, and performs\n\n\nprepare_plugins\nRegisters the plugins for the given configuration.\n\n\n\n\n\ncli.config.check_remote_config(config)\nFirst, determines if the passed config is a valid HTTPS URL. Then, attempts to query\nfor it and parse its content, first as JSON, then as YAML (YAML is preferred).\nFinally, the parsed content is written to a local file and its path is returned.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[str, Path]\nHTTPS URL to a YAML or JSON file.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nUnion[str, Path]\nEither the original config if it’s not a valid HTTPS URL, or the path to the\n\n\n\nUnion[str, Path]\ndownloaded remote config.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the remote configuration is neither valid JSON or YAML.\n\n\n\nRuntimeError\nIf some request-related exception occurs from the file download.\n\n\n\nException\nCatch-all for any other exception.\n\n\n\n\n\n\n\ncli.config.choose_config(path)\nHelper method for choosing a axolotl config YAML file (considering only files\nending with .yml or .yaml). If more than one config file exists in the passed\npath, the user is prompted to choose one.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\npath\nPath\nDirectory in which config file(s) are stored.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr\nPath to either (1) the sole YAML file, or (2) if more than one YAML files exist,\n\n\n\nstr\nthe user-selected YAML file.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf no YAML files are found in the given path.\n\n\n\n\n\n\n\ncli.config.load_cfg(config=Path('examples/'), **kwargs)\nLoads the axolotl configuration stored at config, validates it, and performs\nvarious setup.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr | Path | DictDefault\nPath (local or remote) to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nDictDefault\nDictDefault mapping configuration keys to values.\n\n\n\n\n\n\n\ncli.config.prepare_plugins(cfg)\nRegisters the plugins for the given configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired"
  },
  {
    "objectID": "docs/api/cli.config.html#functions",
    "href": "docs/api/cli.config.html#functions",
    "title": "cli.config",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ncheck_remote_config\nFirst, determines if the passed config is a valid HTTPS URL. Then, attempts to query\n\n\nchoose_config\nHelper method for choosing a axolotl config YAML file (considering only files\n\n\nload_cfg\nLoads the axolotl configuration stored at config, validates it, and performs\n\n\nprepare_plugins\nRegisters the plugins for the given configuration.\n\n\n\n\n\ncli.config.check_remote_config(config)\nFirst, determines if the passed config is a valid HTTPS URL. Then, attempts to query\nfor it and parse its content, first as JSON, then as YAML (YAML is preferred).\nFinally, the parsed content is written to a local file and its path is returned.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[str, Path]\nHTTPS URL to a YAML or JSON file.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nUnion[str, Path]\nEither the original config if it’s not a valid HTTPS URL, or the path to the\n\n\n\nUnion[str, Path]\ndownloaded remote config.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf the remote configuration is neither valid JSON or YAML.\n\n\n\nRuntimeError\nIf some request-related exception occurs from the file download.\n\n\n\nException\nCatch-all for any other exception.\n\n\n\n\n\n\n\ncli.config.choose_config(path)\nHelper method for choosing a axolotl config YAML file (considering only files\nending with .yml or .yaml). If more than one config file exists in the passed\npath, the user is prompted to choose one.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\npath\nPath\nDirectory in which config file(s) are stored.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr\nPath to either (1) the sole YAML file, or (2) if more than one YAML files exist,\n\n\n\nstr\nthe user-selected YAML file.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf no YAML files are found in the given path.\n\n\n\n\n\n\n\ncli.config.load_cfg(config=Path('examples/'), **kwargs)\nLoads the axolotl configuration stored at config, validates it, and performs\nvarious setup.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nstr | Path | DictDefault\nPath (local or remote) to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nDictDefault\nDictDefault mapping configuration keys to values.\n\n\n\n\n\n\n\ncli.config.prepare_plugins(cfg)\nRegisters the plugins for the given configuration.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired"
  },
  {
    "objectID": "docs/api/monkeypatch.gradient_checkpointing.offload_disk.html",
    "href": "docs/api/monkeypatch.gradient_checkpointing.offload_disk.html",
    "title": "monkeypatch.gradient_checkpointing.offload_disk",
    "section": "",
    "text": "monkeypatch.gradient_checkpointing.offload_disk\nDISCO - DIsk-based Storage and Checkpointing with Optimized prefetching\n\n\n\n\n\nName\nDescription\n\n\n\n\nDisco\nDisco: DIsk-based Storage and Checkpointing with Optimized prefetching\n\n\nDiskOffloadManager\nManages offloaded tensors and handles prefetching in a separate thread.\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco()\nDisco: DIsk-based Storage and Checkpointing with Optimized prefetching\nAdvanced disk-based gradient checkpointer with prefetching.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass that loads activations from disk with prefetching\n\n\nforward\nForward pass that offloads activations to disk asynchronously\n\n\nget_instance\nGet or create the offload manager\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco.backward(\n    ctx,\n    *grad_outputs,\n)\nBackward pass that loads activations from disk with prefetching\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco.forward(\n    ctx,\n    forward_function,\n    hidden_states,\n    *args,\n    prefetch_size=1,\n    prefetch_to_gpu=True,\n    save_workers=4,\n)\nForward pass that offloads activations to disk asynchronously\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco.get_instance(\n    prefetch_size=1,\n    prefetch_to_gpu=True,\n    save_workers=4,\n)\nGet or create the offload manager\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager(\n    prefetch_size=3,\n    prefetch_to_gpu=True,\n    save_workers=4,\n)\nManages offloaded tensors and handles prefetching in a separate thread.\nIncludes synchronization to prevent race conditions.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncleanup\nClean up all temp files and stop prefetch thread with proper synchronization\n\n\ncleanup_tensor\nClean up a specific tensor file after it’s been used\n\n\nload_tensor\nLoad tensor from disk or prefetch cache with proper synchronization\n\n\nsave_tensor\nSave tensor to disk asynchronously and return file path with thread-safe operations\n\n\ntrigger_prefetch\nTrigger prefetching of the next N tensors with proper synchronization\n\n\nwait_for_save\nWait for a tensor to be saved to disk\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.cleanup()\nClean up all temp files and stop prefetch thread with proper synchronization\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.cleanup_tensor(\n    file_path,\n)\nClean up a specific tensor file after it’s been used\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.load_tensor(\n    file_path,\n    target_device='cuda',\n)\nLoad tensor from disk or prefetch cache with proper synchronization\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.save_tensor(\n    tensor,\n)\nSave tensor to disk asynchronously and return file path with thread-safe operations\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.trigger_prefetch(\n    n=None,\n)\nTrigger prefetching of the next N tensors with proper synchronization\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.wait_for_save(\n    file_path,\n    timeout=None,\n)\nWait for a tensor to be saved to disk"
  },
  {
    "objectID": "docs/api/monkeypatch.gradient_checkpointing.offload_disk.html#classes",
    "href": "docs/api/monkeypatch.gradient_checkpointing.offload_disk.html#classes",
    "title": "monkeypatch.gradient_checkpointing.offload_disk",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nDisco\nDisco: DIsk-based Storage and Checkpointing with Optimized prefetching\n\n\nDiskOffloadManager\nManages offloaded tensors and handles prefetching in a separate thread.\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco()\nDisco: DIsk-based Storage and Checkpointing with Optimized prefetching\nAdvanced disk-based gradient checkpointer with prefetching.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass that loads activations from disk with prefetching\n\n\nforward\nForward pass that offloads activations to disk asynchronously\n\n\nget_instance\nGet or create the offload manager\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco.backward(\n    ctx,\n    *grad_outputs,\n)\nBackward pass that loads activations from disk with prefetching\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco.forward(\n    ctx,\n    forward_function,\n    hidden_states,\n    *args,\n    prefetch_size=1,\n    prefetch_to_gpu=True,\n    save_workers=4,\n)\nForward pass that offloads activations to disk asynchronously\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.Disco.get_instance(\n    prefetch_size=1,\n    prefetch_to_gpu=True,\n    save_workers=4,\n)\nGet or create the offload manager\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager(\n    prefetch_size=3,\n    prefetch_to_gpu=True,\n    save_workers=4,\n)\nManages offloaded tensors and handles prefetching in a separate thread.\nIncludes synchronization to prevent race conditions.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncleanup\nClean up all temp files and stop prefetch thread with proper synchronization\n\n\ncleanup_tensor\nClean up a specific tensor file after it’s been used\n\n\nload_tensor\nLoad tensor from disk or prefetch cache with proper synchronization\n\n\nsave_tensor\nSave tensor to disk asynchronously and return file path with thread-safe operations\n\n\ntrigger_prefetch\nTrigger prefetching of the next N tensors with proper synchronization\n\n\nwait_for_save\nWait for a tensor to be saved to disk\n\n\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.cleanup()\nClean up all temp files and stop prefetch thread with proper synchronization\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.cleanup_tensor(\n    file_path,\n)\nClean up a specific tensor file after it’s been used\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.load_tensor(\n    file_path,\n    target_device='cuda',\n)\nLoad tensor from disk or prefetch cache with proper synchronization\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.save_tensor(\n    tensor,\n)\nSave tensor to disk asynchronously and return file path with thread-safe operations\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.trigger_prefetch(\n    n=None,\n)\nTrigger prefetching of the next N tensors with proper synchronization\n\n\n\nmonkeypatch.gradient_checkpointing.offload_disk.DiskOffloadManager.wait_for_save(\n    file_path,\n    timeout=None,\n)\nWait for a tensor to be saved to disk"
  },
  {
    "objectID": "docs/api/monkeypatch.stablelm_attn_hijack_flash.html",
    "href": "docs/api/monkeypatch.stablelm_attn_hijack_flash.html",
    "title": "monkeypatch.stablelm_attn_hijack_flash",
    "section": "",
    "text": "monkeypatch.stablelm_attn_hijack_flash\nPyTorch StableLM Epoch model.\n\n\n\n\n\nName\nDescription\n\n\n\n\nrepeat_kv\nThis is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,\n\n\nrotate_half\nRotates half the hidden dims of the input.\n\n\n\n\n\nmonkeypatch.stablelm_attn_hijack_flash.repeat_kv(hidden_states, n_rep)\nThis is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,\nnum_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)\n\n\n\nmonkeypatch.stablelm_attn_hijack_flash.rotate_half(x)\nRotates half the hidden dims of the input."
  },
  {
    "objectID": "docs/api/monkeypatch.stablelm_attn_hijack_flash.html#functions",
    "href": "docs/api/monkeypatch.stablelm_attn_hijack_flash.html#functions",
    "title": "monkeypatch.stablelm_attn_hijack_flash",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nrepeat_kv\nThis is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,\n\n\nrotate_half\nRotates half the hidden dims of the input.\n\n\n\n\n\nmonkeypatch.stablelm_attn_hijack_flash.repeat_kv(hidden_states, n_rep)\nThis is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,\nnum_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)\n\n\n\nmonkeypatch.stablelm_attn_hijack_flash.rotate_half(x)\nRotates half the hidden dims of the input."
  },
  {
    "objectID": "docs/api/kernels.lora.html",
    "href": "docs/api/kernels.lora.html",
    "title": "kernels.lora",
    "section": "",
    "text": "kernels.lora\nModule for definition of Low-Rank Adaptation (LoRA) Triton kernels.\nSee “LoRA: Low-Rank Adaptation of Large Language Models”\n(https://arxiv.org/abs/2106.09685).\nCredit to unsloth (https://unsloth.ai/) for inspiration for this implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nLoRA_MLP\nOptimized LoRA MLP implementation.\n\n\nLoRA_O\nOptimized LoRA implementation for output projection.\n\n\nLoRA_QKV\nOptimized LoRA QKV implementation with quantization support.\n\n\n\n\n\nkernels.lora.LoRA_MLP()\nOptimized LoRA MLP implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nPerforms backward pass computation for LoRA MLP.\n\n\nforward\nForward pass for LoRA MLP.\n\n\n\n\n\nkernels.lora.LoRA_MLP.backward(ctx, grad_output)\nPerforms backward pass computation for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nContext object storing tensors saved during forward pass\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to layer output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor | None\nTuple containing gradients for all inputs from forward pass:\n\n\n\nNone\n- Input gradient tensor (or None)\n\n\n\nNone\n- None for weights/quantization states\n\n\n\ntorch.Tensor | None\n- LoRA A/B matrix gradients (or None)\n\n\n\ntorch.Tensor | None\n- None for scaling factors\n\n\n\nNone\n- None for activation functions and flags\n\n\n\n\n\n\n\nkernels.lora.LoRA_MLP.forward(\n    ctx,\n    X,\n    gate_weight,\n    gate_quant,\n    gate_A,\n    gate_B,\n    gate_scale,\n    up_weight,\n    up_quant,\n    up_A,\n    up_B,\n    up_scale,\n    down_weight,\n    down_quant,\n    down_A,\n    down_B,\n    down_scale,\n    activation_fn,\n    activation_fn_backward,\n    inplace=True,\n)\nForward pass for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\n\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput features\nrequired\n\n\ngate_weight\ntorch.Tensor\nGate projection weight\nrequired\n\n\ngate_quant\nobject | None\nGate quantization state\nrequired\n\n\ngate_A\ntorch.Tensor | None\nGate LoRA A matrix\nrequired\n\n\ngate_B\ntorch.Tensor | None\nGate LoRA B matrix\nrequired\n\n\ngate_scale\nfloat\nGate LoRA scale\nrequired\n\n\nup_weight\ntorch.Tensor\nUp-projection weight\nrequired\n\n\nup_quant\nobject | None\nUp-projection quantization state\nrequired\n\n\nup_A\ntorch.Tensor | None\nUp-projection LoRA A matrix\nrequired\n\n\nup_B\ntorch.Tensor | None\nUp-projection LoRA B matrix\nrequired\n\n\nup_scale\nfloat\nUp-projection LoRA scale\nrequired\n\n\ndown_weight\ntorch.Tensor\nDown-projection weight\nrequired\n\n\ndown_quant\nobject | None\nDown-projection quantization state\nrequired\n\n\ndown_A\ntorch.Tensor | None\nDown-projection LoRA A matrix\nrequired\n\n\ndown_B\ntorch.Tensor | None\nDown-projection LoRA B matrix\nrequired\n\n\ndown_scale\nfloat\nDown-projection LoRA scale\nrequired\n\n\nactivation_fn\nCallable\nForward activation function\nrequired\n\n\nactivation_fn_backward\nCallable\nBackward activation function\nrequired\n\n\ninplace\nbool | None\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput transformed by multi-layer perceptron and activation function\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_O()\nOptimized LoRA implementation for output projection.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA output projection.\n\n\nforward\nForward pass for output projection with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_O.backward(ctx, dY)\nBackward pass computing gradients for LoRA output projection.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\ndY\ntorch.Tensor\nGradient of loss with respect to output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None, None, torch.Tensor | None, torch.Tensor | None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_O.forward(ctx, X, W, W_quant, A, B, S)\nForward pass for output projection with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nW\ntorch.Tensor\nOutput projection weight\nrequired\n\n\nW_quant\nQuantState | None\nWeight quantization state\nrequired\n\n\nA\ntorch.Tensor | None\nLoRA A matrix\nrequired\n\n\nB\ntorch.Tensor | None\nLoRA B matrix\nrequired\n\n\nS\nfloat\nLoRA scaling factor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput projection tensor\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV()\nOptimized LoRA QKV implementation with quantization support.\nImplements efficient computation of query, key, value projections with LoRA,\nsupporting quantization and memory optimization.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA QKV.\n\n\nforward\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_QKV.backward(ctx, q_grad, k_grad, v_grad)\nBackward pass computing gradients for LoRA QKV.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nq_grad\ntorch.Tensor\nGradient for query projection\nrequired\n\n\nk_grad\ntorch.Tensor\nGradient for key projection\nrequired\n\n\nv_grad\ntorch.Tensor\nGradient for value projection\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV.forward(\n    ctx,\n    X,\n    q_weight,\n    q_quant,\n    q_A,\n    q_B,\n    q_scale,\n    k_weight,\n    k_quant,\n    k_A,\n    k_B,\n    k_scale,\n    v_weight,\n    v_quant,\n    v_A,\n    v_B,\n    v_scale,\n    inplace=True,\n)\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nq_weight\ntorch.Tensor\nQuery projection weight\nrequired\n\n\nq_quant\nQuantState | None\nQuery quantization state\nrequired\n\n\nq_A\ntorch.Tensor | None\nQuery LoRA A matrix\nrequired\n\n\nq_B\ntorch.Tensor | None\nQuery LoRA B matrix\nrequired\n\n\nq_scale\nfloat\nQuery LoRA scale\nrequired\n\n\nk_weight\ntorch.Tensor\nKey projection weight\nrequired\n\n\nk_quant\nQuantState | None\nKey quantization state\nrequired\n\n\nk_A\ntorch.Tensor | None\nKey LoRA A matrix\nrequired\n\n\nk_B\ntorch.Tensor | None\nKey LoRA B matrix\nrequired\n\n\nk_scale\nfloat\nKey LoRA scale\nrequired\n\n\nv_weight\ntorch.Tensor\nValue projection weight\nrequired\n\n\nv_quant\nQuantState | None\nValue quantization state\nrequired\n\n\nv_A\ntorch.Tensor | None\nValue LoRA A matrix\nrequired\n\n\nv_B\ntorch.Tensor | None\nValue LoRA B matrix\nrequired\n\n\nv_scale\nfloat\nValue LoRA scale\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_lora_mlp_geglu\nApplies LoRA to MLP layer with GEGLU activation.\n\n\napply_lora_mlp_swiglu\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\napply_lora_o\nApplies LoRA to output projection layer.\n\n\napply_lora_qkv\nApplies LoRA to compute Query, Key, Value projections.\n\n\nget_lora_parameters\nGets LoRA parameters from a projection module.\n\n\nmatmul_lora\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\nkernels.lora.apply_lora_mlp_geglu(self, X, inplace=True)\nApplies LoRA to MLP layer with GEGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with GEGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_mlp_swiglu(self, X, inplace=True)\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with SwiGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_o(self, X)\nApplies LoRA to output projection layer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTransformed output tensor\n\n\n\n\n\n\n\nkernels.lora.apply_lora_qkv(self, X, inplace=True)\nApplies LoRA to compute Query, Key, Value projections.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\nkernels.lora.get_lora_parameters(proj)\nGets LoRA parameters from a projection module.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nproj\nnn.Module\nThe projection module to extract parameters from.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nA tuple containing the base weight matrix, quantization state, LoRA A matrix,\n\n\n\nQuantState | None\nLoRA B matrix, and scaling factor. States and matrices may be None if not\n\n\n\ntorch.Tensor | None\navailable.\n\n\n\n\n\n\n\nkernels.lora.matmul_lora(X, W, W_quant, A, B, s, out=None)\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor [*, in_features]\nrequired\n\n\nW\ntorch.Tensor\nBase weight matrix [out_features, in_features]\nrequired\n\n\nW_quant\nQuantState\nQuantization state for W\nrequired\n\n\nA\ntorch.Tensor\nLoRA A matrix [rank, in_features]\nrequired\n\n\nB\ntorch.Tensor\nLoRA B matrix [out_features, rank]\nrequired\n\n\ns\nfloat\nLoRA scaling factor\nrequired\n\n\nout\ntorch.Tensor | None\nOptional output tensor for inplace operations\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nResult of X @ W + X @ A @ B"
  },
  {
    "objectID": "docs/api/kernels.lora.html#classes",
    "href": "docs/api/kernels.lora.html#classes",
    "title": "kernels.lora",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nLoRA_MLP\nOptimized LoRA MLP implementation.\n\n\nLoRA_O\nOptimized LoRA implementation for output projection.\n\n\nLoRA_QKV\nOptimized LoRA QKV implementation with quantization support.\n\n\n\n\n\nkernels.lora.LoRA_MLP()\nOptimized LoRA MLP implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nPerforms backward pass computation for LoRA MLP.\n\n\nforward\nForward pass for LoRA MLP.\n\n\n\n\n\nkernels.lora.LoRA_MLP.backward(ctx, grad_output)\nPerforms backward pass computation for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nContext object storing tensors saved during forward pass\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to layer output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor | None\nTuple containing gradients for all inputs from forward pass:\n\n\n\nNone\n- Input gradient tensor (or None)\n\n\n\nNone\n- None for weights/quantization states\n\n\n\ntorch.Tensor | None\n- LoRA A/B matrix gradients (or None)\n\n\n\ntorch.Tensor | None\n- None for scaling factors\n\n\n\nNone\n- None for activation functions and flags\n\n\n\n\n\n\n\nkernels.lora.LoRA_MLP.forward(\n    ctx,\n    X,\n    gate_weight,\n    gate_quant,\n    gate_A,\n    gate_B,\n    gate_scale,\n    up_weight,\n    up_quant,\n    up_A,\n    up_B,\n    up_scale,\n    down_weight,\n    down_quant,\n    down_A,\n    down_B,\n    down_scale,\n    activation_fn,\n    activation_fn_backward,\n    inplace=True,\n)\nForward pass for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\n\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput features\nrequired\n\n\ngate_weight\ntorch.Tensor\nGate projection weight\nrequired\n\n\ngate_quant\nobject | None\nGate quantization state\nrequired\n\n\ngate_A\ntorch.Tensor | None\nGate LoRA A matrix\nrequired\n\n\ngate_B\ntorch.Tensor | None\nGate LoRA B matrix\nrequired\n\n\ngate_scale\nfloat\nGate LoRA scale\nrequired\n\n\nup_weight\ntorch.Tensor\nUp-projection weight\nrequired\n\n\nup_quant\nobject | None\nUp-projection quantization state\nrequired\n\n\nup_A\ntorch.Tensor | None\nUp-projection LoRA A matrix\nrequired\n\n\nup_B\ntorch.Tensor | None\nUp-projection LoRA B matrix\nrequired\n\n\nup_scale\nfloat\nUp-projection LoRA scale\nrequired\n\n\ndown_weight\ntorch.Tensor\nDown-projection weight\nrequired\n\n\ndown_quant\nobject | None\nDown-projection quantization state\nrequired\n\n\ndown_A\ntorch.Tensor | None\nDown-projection LoRA A matrix\nrequired\n\n\ndown_B\ntorch.Tensor | None\nDown-projection LoRA B matrix\nrequired\n\n\ndown_scale\nfloat\nDown-projection LoRA scale\nrequired\n\n\nactivation_fn\nCallable\nForward activation function\nrequired\n\n\nactivation_fn_backward\nCallable\nBackward activation function\nrequired\n\n\ninplace\nbool | None\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput transformed by multi-layer perceptron and activation function\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_O()\nOptimized LoRA implementation for output projection.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA output projection.\n\n\nforward\nForward pass for output projection with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_O.backward(ctx, dY)\nBackward pass computing gradients for LoRA output projection.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\ndY\ntorch.Tensor\nGradient of loss with respect to output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None, None, torch.Tensor | None, torch.Tensor | None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_O.forward(ctx, X, W, W_quant, A, B, S)\nForward pass for output projection with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nW\ntorch.Tensor\nOutput projection weight\nrequired\n\n\nW_quant\nQuantState | None\nWeight quantization state\nrequired\n\n\nA\ntorch.Tensor | None\nLoRA A matrix\nrequired\n\n\nB\ntorch.Tensor | None\nLoRA B matrix\nrequired\n\n\nS\nfloat\nLoRA scaling factor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput projection tensor\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV()\nOptimized LoRA QKV implementation with quantization support.\nImplements efficient computation of query, key, value projections with LoRA,\nsupporting quantization and memory optimization.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA QKV.\n\n\nforward\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_QKV.backward(ctx, q_grad, k_grad, v_grad)\nBackward pass computing gradients for LoRA QKV.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nq_grad\ntorch.Tensor\nGradient for query projection\nrequired\n\n\nk_grad\ntorch.Tensor\nGradient for key projection\nrequired\n\n\nv_grad\ntorch.Tensor\nGradient for value projection\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV.forward(\n    ctx,\n    X,\n    q_weight,\n    q_quant,\n    q_A,\n    q_B,\n    q_scale,\n    k_weight,\n    k_quant,\n    k_A,\n    k_B,\n    k_scale,\n    v_weight,\n    v_quant,\n    v_A,\n    v_B,\n    v_scale,\n    inplace=True,\n)\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nq_weight\ntorch.Tensor\nQuery projection weight\nrequired\n\n\nq_quant\nQuantState | None\nQuery quantization state\nrequired\n\n\nq_A\ntorch.Tensor | None\nQuery LoRA A matrix\nrequired\n\n\nq_B\ntorch.Tensor | None\nQuery LoRA B matrix\nrequired\n\n\nq_scale\nfloat\nQuery LoRA scale\nrequired\n\n\nk_weight\ntorch.Tensor\nKey projection weight\nrequired\n\n\nk_quant\nQuantState | None\nKey quantization state\nrequired\n\n\nk_A\ntorch.Tensor | None\nKey LoRA A matrix\nrequired\n\n\nk_B\ntorch.Tensor | None\nKey LoRA B matrix\nrequired\n\n\nk_scale\nfloat\nKey LoRA scale\nrequired\n\n\nv_weight\ntorch.Tensor\nValue projection weight\nrequired\n\n\nv_quant\nQuantState | None\nValue quantization state\nrequired\n\n\nv_A\ntorch.Tensor | None\nValue LoRA A matrix\nrequired\n\n\nv_B\ntorch.Tensor | None\nValue LoRA B matrix\nrequired\n\n\nv_scale\nfloat\nValue LoRA scale\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple of (Query, Key, Value) projection tensors"
  },
  {
    "objectID": "docs/api/kernels.lora.html#functions",
    "href": "docs/api/kernels.lora.html#functions",
    "title": "kernels.lora",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\napply_lora_mlp_geglu\nApplies LoRA to MLP layer with GEGLU activation.\n\n\napply_lora_mlp_swiglu\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\napply_lora_o\nApplies LoRA to output projection layer.\n\n\napply_lora_qkv\nApplies LoRA to compute Query, Key, Value projections.\n\n\nget_lora_parameters\nGets LoRA parameters from a projection module.\n\n\nmatmul_lora\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\nkernels.lora.apply_lora_mlp_geglu(self, X, inplace=True)\nApplies LoRA to MLP layer with GEGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with GEGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_mlp_swiglu(self, X, inplace=True)\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with SwiGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_o(self, X)\nApplies LoRA to output projection layer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTransformed output tensor\n\n\n\n\n\n\n\nkernels.lora.apply_lora_qkv(self, X, inplace=True)\nApplies LoRA to compute Query, Key, Value projections.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\nkernels.lora.get_lora_parameters(proj)\nGets LoRA parameters from a projection module.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nproj\nnn.Module\nThe projection module to extract parameters from.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nA tuple containing the base weight matrix, quantization state, LoRA A matrix,\n\n\n\nQuantState | None\nLoRA B matrix, and scaling factor. States and matrices may be None if not\n\n\n\ntorch.Tensor | None\navailable.\n\n\n\n\n\n\n\nkernels.lora.matmul_lora(X, W, W_quant, A, B, s, out=None)\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor [*, in_features]\nrequired\n\n\nW\ntorch.Tensor\nBase weight matrix [out_features, in_features]\nrequired\n\n\nW_quant\nQuantState\nQuantization state for W\nrequired\n\n\nA\ntorch.Tensor\nLoRA A matrix [rank, in_features]\nrequired\n\n\nB\ntorch.Tensor\nLoRA B matrix [out_features, rank]\nrequired\n\n\ns\nfloat\nLoRA scaling factor\nrequired\n\n\nout\ntorch.Tensor | None\nOptional output tensor for inplace operations\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nResult of X @ W + X @ A @ B"
  },
  {
    "objectID": "docs/api/prompt_strategies.base.html",
    "href": "docs/api/prompt_strategies.base.html",
    "title": "prompt_strategies.base",
    "section": "",
    "text": "prompt_strategies.base\nprompt_strategies.base\nmodule for base dataset transform strategies"
  },
  {
    "objectID": "docs/api/core.trainers.relora.html",
    "href": "docs/api/core.trainers.relora.html",
    "title": "core.trainers.relora",
    "section": "",
    "text": "core.trainers.relora\nModule for ReLoRA trainer\n\n\n\n\n\nName\nDescription\n\n\n\n\nReLoRATrainer\nTrainer subclass that uses the OneCycleLR scheduler\n\n\n\n\n\ncore.trainers.relora.ReLoRATrainer(*args, **kwargs)\nTrainer subclass that uses the OneCycleLR scheduler"
  },
  {
    "objectID": "docs/api/core.trainers.relora.html#classes",
    "href": "docs/api/core.trainers.relora.html#classes",
    "title": "core.trainers.relora",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nReLoRATrainer\nTrainer subclass that uses the OneCycleLR scheduler\n\n\n\n\n\ncore.trainers.relora.ReLoRATrainer(*args, **kwargs)\nTrainer subclass that uses the OneCycleLR scheduler"
  },
  {
    "objectID": "docs/api/loaders.model.html",
    "href": "docs/api/loaders.model.html",
    "title": "loaders.model",
    "section": "",
    "text": "loaders.model\nModel loader class implementation for loading, configuring, and patching various\nmodels.\n\n\n\n\n\nName\nDescription\n\n\n\n\nModelLoader\nManages model configuration, initialization and application of patches during\n\n\n\n\n\nloaders.model.ModelLoader(\n    cfg,\n    tokenizer,\n    *,\n    inference=False,\n    reference_model=False,\n    **kwargs,\n)\nManages model configuration, initialization and application of patches during\nmodel loading.\nThis class orchestrates the entire process of loading a model from configuration to\nfinal preparation. It handles device mapping, quantization, attention mechanisms,\nadapter integration, and various optimizations.\n\n\n\nLoading and validating model configuration\nApplying monkey patches for optimizations / fixes\nSetting up device mapping (including multi-GPU configurations)\nConfiguring quantization\nSetting attention mechanisms (Flash Attention, SDPA, etc.)\nLoading and initializing the model\nApplying adapters (LoRA, QLoRA, etc.)\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nmodel\nPreTrainedModel | PeftModel | PeftMixedModel\nThe loaded model instance (available after load() is called).\n\n\nmodel_kwargs\ndict[str, Any]\nDictionary of keyword arguments passed to model initialization.\n\n\nbase_model\n\nName or path of the base model to load.\n\n\nmodel_type\n\nType of model to load (e.g., AutoModelForCausalLM).\n\n\nmodel_config\n\nConfiguration object for the model.\n\n\nauto_model_loader\n\nclass used for loading the model (default: AutoModelForCausalLM).\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nload\nLoad and prepare the model with all configurations and patches.\n\n\n\n\n\nloaders.model.ModelLoader.load()\nLoad and prepare the model with all configurations and patches.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PreTrainedModel | PeftModelForCausalLM, PeftConfig | None]\nA tuple with the loaded model and its LoRA configuration (if applicable)."
  },
  {
    "objectID": "docs/api/loaders.model.html#classes",
    "href": "docs/api/loaders.model.html#classes",
    "title": "loaders.model",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nModelLoader\nManages model configuration, initialization and application of patches during\n\n\n\n\n\nloaders.model.ModelLoader(\n    cfg,\n    tokenizer,\n    *,\n    inference=False,\n    reference_model=False,\n    **kwargs,\n)\nManages model configuration, initialization and application of patches during\nmodel loading.\nThis class orchestrates the entire process of loading a model from configuration to\nfinal preparation. It handles device mapping, quantization, attention mechanisms,\nadapter integration, and various optimizations.\n\n\n\nLoading and validating model configuration\nApplying monkey patches for optimizations / fixes\nSetting up device mapping (including multi-GPU configurations)\nConfiguring quantization\nSetting attention mechanisms (Flash Attention, SDPA, etc.)\nLoading and initializing the model\nApplying adapters (LoRA, QLoRA, etc.)\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nmodel\nPreTrainedModel | PeftModel | PeftMixedModel\nThe loaded model instance (available after load() is called).\n\n\nmodel_kwargs\ndict[str, Any]\nDictionary of keyword arguments passed to model initialization.\n\n\nbase_model\n\nName or path of the base model to load.\n\n\nmodel_type\n\nType of model to load (e.g., AutoModelForCausalLM).\n\n\nmodel_config\n\nConfiguration object for the model.\n\n\nauto_model_loader\n\nclass used for loading the model (default: AutoModelForCausalLM).\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\nload\nLoad and prepare the model with all configurations and patches.\n\n\n\n\n\nloaders.model.ModelLoader.load()\nLoad and prepare the model with all configurations and patches.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[PreTrainedModel | PeftModelForCausalLM, PeftConfig | None]\nA tuple with the loaded model and its LoRA configuration (if applicable)."
  },
  {
    "objectID": "docs/api/utils.callbacks.perplexity.html",
    "href": "docs/api/utils.callbacks.perplexity.html",
    "title": "utils.callbacks.perplexity",
    "section": "",
    "text": "utils.callbacks.perplexity\ncallback to calculate perplexity as an evaluation metric.\n\n\n\n\n\nName\nDescription\n\n\n\n\nPerplexity\nCalculate perplexity as defined in https://huggingface.co/docs/transformers/en/perplexity.\n\n\n\n\n\nutils.callbacks.perplexity.Perplexity(tokenizer, max_seq_len, stride=512)\nCalculate perplexity as defined in https://huggingface.co/docs/transformers/en/perplexity.\nThis is a custom variant that doesn’t re-tokenize the input or re-load the model.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncompute\nCompute perplexity in a fixed length sliding window across the sequence.\n\n\n\n\n\nutils.callbacks.perplexity.Perplexity.compute(model, references=None)\nCompute perplexity in a fixed length sliding window across the sequence."
  },
  {
    "objectID": "docs/api/utils.callbacks.perplexity.html#classes",
    "href": "docs/api/utils.callbacks.perplexity.html#classes",
    "title": "utils.callbacks.perplexity",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\nPerplexity\nCalculate perplexity as defined in https://huggingface.co/docs/transformers/en/perplexity.\n\n\n\n\n\nutils.callbacks.perplexity.Perplexity(tokenizer, max_seq_len, stride=512)\nCalculate perplexity as defined in https://huggingface.co/docs/transformers/en/perplexity.\nThis is a custom variant that doesn’t re-tokenize the input or re-load the model.\n\n\n\n\n\nName\nDescription\n\n\n\n\ncompute\nCompute perplexity in a fixed length sliding window across the sequence.\n\n\n\n\n\nutils.callbacks.perplexity.Perplexity.compute(model, references=None)\nCompute perplexity in a fixed length sliding window across the sequence."
  },
  {
    "objectID": "docs/api/cli.inference.html",
    "href": "docs/api/cli.inference.html",
    "title": "cli.inference",
    "section": "",
    "text": "cli.inference\nCLI to run inference on a trained model.\n\n\n\n\n\nName\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_inference or do_inference_gradio.\n\n\ndo_inference\nRuns inference on the command line in a loop. User input is accepted, a chat template\n\n\ndo_inference_gradio\nRuns inference in a Gradio interface. User input is accepted, a chat template is\n\n\nget_multi_line_input\nGets multi-line input from terminal.\n\n\n\n\n\ncli.inference.do_cli(config=Path('examples/'), gradio=False, **kwargs)\nParses axolotl config, CLI args, and calls do_inference or do_inference_gradio.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.inference.do_inference(cfg, cli_args)\nRuns inference on the command line in a loop. User input is accepted, a chat template\nis (optionally) applied, and the model specified in the axolotl config is used to\ngenerate completions according to a default generation config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nInferenceCliArgs\nInference-specific CLI arguments.\nrequired\n\n\n\n\n\n\n\ncli.inference.do_inference_gradio(cfg, cli_args)\nRuns inference in a Gradio interface. User input is accepted, a chat template is\n(optionally) applied, and the model specified in the axolotl config is used to\ngenerate completions according to a default generation config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nInferenceCliArgs\nInference-specific CLI arguments.\nrequired\n\n\n\n\n\n\n\ncli.inference.get_multi_line_input()\nGets multi-line input from terminal.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr\nPossibly multi-line, possibly empty stdin input as a string."
  },
  {
    "objectID": "docs/api/cli.inference.html#functions",
    "href": "docs/api/cli.inference.html#functions",
    "title": "cli.inference",
    "section": "",
    "text": "Name\nDescription\n\n\n\n\ndo_cli\nParses axolotl config, CLI args, and calls do_inference or do_inference_gradio.\n\n\ndo_inference\nRuns inference on the command line in a loop. User input is accepted, a chat template\n\n\ndo_inference_gradio\nRuns inference in a Gradio interface. User input is accepted, a chat template is\n\n\nget_multi_line_input\nGets multi-line input from terminal.\n\n\n\n\n\ncli.inference.do_cli(config=Path('examples/'), gradio=False, **kwargs)\nParses axolotl config, CLI args, and calls do_inference or do_inference_gradio.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconfig\nUnion[Path, str]\nPath to axolotl config YAML file.\nPath('examples/')\n\n\nkwargs\n\nAdditional keyword arguments to override config file values.\n{}\n\n\n\n\n\n\n\ncli.inference.do_inference(cfg, cli_args)\nRuns inference on the command line in a loop. User input is accepted, a chat template\nis (optionally) applied, and the model specified in the axolotl config is used to\ngenerate completions according to a default generation config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nInferenceCliArgs\nInference-specific CLI arguments.\nrequired\n\n\n\n\n\n\n\ncli.inference.do_inference_gradio(cfg, cli_args)\nRuns inference in a Gradio interface. User input is accepted, a chat template is\n(optionally) applied, and the model specified in the axolotl config is used to\ngenerate completions according to a default generation config.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\ncli_args\nInferenceCliArgs\nInference-specific CLI arguments.\nrequired\n\n\n\n\n\n\n\ncli.inference.get_multi_line_input()\nGets multi-line input from terminal.\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nstr\nPossibly multi-line, possibly empty stdin input as a string."
  },
  {
    "objectID": "docs/batch_vs_grad.html",
    "href": "docs/batch_vs_grad.html",
    "title": "Batch size vs Gradient accumulation",
    "section": "",
    "text": "Gradient accumulation means accumulating gradients over several mini-batches and updating the model weights afterward. When the samples in each batch are diverse, this technique doesn’t significantly impact learning.\nThis method allows for effective training with larger effective batch sizes without needing proportionally larger memory. Here’s why:\n\nMemory Consumption with Batch Size: The primary reason increasing the batch size impacts memory is due to the storage requirements for intermediate activations. When you forward propagate a batch through a network, you have to store the activations at each layer for each sample in the batch, because these activations are used during backpropagation to compute gradients. Therefore, larger batches mean more activations, leading to greater GPU memory consumption.\nGradient Accumulation: With gradient accumulation, you’re effectively simulating a larger batch size by accumulating gradients over several smaller batches (or micro-batches). However, at any given time, you’re only forward and backward propagating a micro-batch. This means you only store activations for the micro-batch, not the full accumulated batch. As a result, you can simulate the effect of a larger batch size without the memory cost of storing activations for a large batch.\n\nExample 1:\nMicro batch size: 3\nGradient accumulation steps: 2\nNumber of GPUs: 3\nTotal batch size = 3 * 2 * 3 = 18\n| GPU 1          | GPU 2          | GPU 3          |\n|----------------|----------------|----------------|\n| S1, S2, S3     | S4, S5, S6     | S7, S8, S9     |\n| e1, e2, e3     | e4, e5, e6     | e7, e8, e9     |\n|----------------|----------------|----------------|\n| → (accumulate) | → (accumulate) | → (accumulate) |\n|----------------|----------------|----------------|\n| S10, S11, S12  | S13, S14, S15  | S16, S17, S18  |\n| e10, e11, e12  | e13, e14, e15  | e16, e17, e18  |\n|----------------|----------------|----------------|\n| → (apply)      | → (apply)      | → (apply)      |\n\nAccumulated gradient for the weight w1 after the second iteration (considering all GPUs):\nTotal gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 + e9 + e10 + e11 + e12 + e13 + e14 + e15 + e16 + e17 + e18\n\nWeight update for w1:\nw1_new = w1_old - learning rate x (Total gradient for w1 / 18)\nExample 2:\nMicro batch size: 2\nGradient accumulation steps: 1\nNumber of GPUs: 3\nTotal batch size = 2 * 1 * 3 = 6\n| GPU 1     | GPU 2     | GPU 3     |\n|-----------|-----------|-----------|\n| S1, S2    | S3, S4    | S5, S6    |\n| e1, e2    | e3, e4    | e5, e6    |\n|-----------|-----------|-----------|\n| → (apply) | → (apply) | → (apply) |\n\nAccumulated gradient for the weight w1 (considering all GPUs):\nTotal gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6\n\nWeight update for w1:\nw1_new = w1_old - learning rate × (Total gradient for w1 / 6)",
    "crumbs": [
      "Core Concepts",
      "Batch size vs Gradient accumulation"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html",
    "href": "docs/sequence_parallelism.html",
    "title": "Sequence Parallelism",
    "section": "",
    "text": "Sequence parallelism is a technique that splits sequences across multiple GPUs,\nallowing you to train with very long sequences that wouldn’t fit on a single GPU. Each\nGPU processes a different portion of the sequence, and the results are aggregated\nthrough a ring communication pattern.",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#when-to-use-sequence-parallelism",
    "href": "docs/sequence_parallelism.html#when-to-use-sequence-parallelism",
    "title": "Sequence Parallelism",
    "section": "When to Use Sequence Parallelism",
    "text": "When to Use Sequence Parallelism\nUse sequence parallelism when:\n\nYou need to train with sequence lengths that don’t fit into a single GPU’s memory\nYou have multiple GPUs available\nYou’re experiencing OOM (Out Of Memory) errors with long sequences",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#configuration",
    "href": "docs/sequence_parallelism.html#configuration",
    "title": "Sequence Parallelism",
    "section": "Configuration",
    "text": "Configuration\nTo enable sequence parallelism, add the following to your configuration file:\n# Set to a divisor (> 1) of the number of GPUs available\nsequence_parallel_degree: 4  # Split sequences across 4 GPUs\n# Optional; strides across the key dimension. Larger values use more memory but should make training faster.\nheads_k_stride: 1\n# Optional; one of \"varlen_llama3\" or \"batch_ring\". Defaults to\n# \"varlen_llama3\" when `sample_packing: true`, and \"batch_ring\" otherwise.\nring_attn_func:\nThe sequence_parallel_degree should be a divisor of the total number of GPUs. For example:\n\nWith 8 GPUs, valid values would be 2, 4, or 8\nWith 4 GPUs, valid values would be 2 or 4",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#implementation-details",
    "href": "docs/sequence_parallelism.html#implementation-details",
    "title": "Sequence Parallelism",
    "section": "Implementation Details",
    "text": "Implementation Details\nWhen sequence parallelism is enabled:\n\nEach sequence is divided into equal chunks across the GPUs in a sequence parallel group\nThe data collator handles the chunking of input_ids, attention_mask, labels, and position_ids\nPosition IDs are adjusted to maintain proper relative positions\nThe trainer uses special ring communication patterns for attention operations",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#requirements",
    "href": "docs/sequence_parallelism.html#requirements",
    "title": "Sequence Parallelism",
    "section": "Requirements",
    "text": "Requirements\nTo use sequence parallelism, you need:\n\nMultiple GPUs (at least 2)\nThe ring-flash-attn package. Install with:\n\npip install axolotl[ring-flash-attn] (preferred)\npip install ring-flash-attn>=0.1.4",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#limitations",
    "href": "docs/sequence_parallelism.html#limitations",
    "title": "Sequence Parallelism",
    "section": "Limitations",
    "text": "Limitations\n\nFlash attention must be enabled for this to work (flash_attention: true in config YAML)\nMay have a small performance overhead due to communication between GPUs",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#example",
    "href": "docs/sequence_parallelism.html#example",
    "title": "Sequence Parallelism",
    "section": "Example",
    "text": "Example\nbase_model: meta-llama/Llama-3-8B-Instruct\nsequence_len: 8192\n\n...\n\nsequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU\n# Optional; strides across the key dimension. Larger values use more memory but should make training faster.\nheads_k_stride: 1\n# Optional; one of \"varlen_llama3\" or \"batch_ring\". Defaults to\n# \"varlen_llama3\" when `sample_packing: true`, and \"batch_ring\" otherwise.\nring_attn_func:\n\n...\nThis will train the Llama 3 8B model with 8K context length, with each sequence split\ninto 2 subsequences of length 4096 across 2 GPUs.",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#sample-packing-with-sequence-parallelism",
    "href": "docs/sequence_parallelism.html#sample-packing-with-sequence-parallelism",
    "title": "Sequence Parallelism",
    "section": "Sample Packing with Sequence Parallelism",
    "text": "Sample Packing with Sequence Parallelism\nSequence parallelism is compatible with Axolotl’s sample packing functionality. When using both features together:\n\nSamples are first packed together\nThe packed sequences are then divided across GPUs in the sequence parallel group\nPosition IDs are automatically adjusted to maintain proper relative positions",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/sequence_parallelism.html#effect-on-batch-size",
    "href": "docs/sequence_parallelism.html#effect-on-batch-size",
    "title": "Sequence Parallelism",
    "section": "Effect on Batch Size",
    "text": "Effect on Batch Size\nWhen using sequence parallelism, your effective global batch size is divided by the sequence_parallel_degree. This happens because:\n\nEach group of sequence_parallel_degree GPUs works on the same batch (just different parts of each sequence)\nThe number of batches processed per step decreases\n\nFor example:\n- With 8 GPUs and no sequence parallelism: 8 different batches processed per step\n- With 8 GPUs and sequence_parallel_degree=4: Only 2 different batches processed per step (each split across 4 GPUs)\n- If your per-GPU micro_batch_size is 2, the global batch size decreases from 16 to 4",
    "crumbs": [
      "Advanced Features",
      "Sequence Parallelism"
    ]
  },
  {
    "objectID": "docs/getting-started.html",
    "href": "docs/getting-started.html",
    "title": "Quickstart",
    "section": "",
    "text": "This guide will walk you through your first model fine-tuning project with Axolotl.",
    "crumbs": [
      "Getting Started",
      "Quickstart"
    ]
  },
  {
    "objectID": "docs/getting-started.html#sec-quick-example",
    "href": "docs/getting-started.html#sec-quick-example",
    "title": "Quickstart",
    "section": "1 Quick Example",
    "text": "1 Quick Example\nLet’s start by fine-tuning a small language model using LoRA. This example uses a 1B parameter model to ensure it runs on most GPUs.\nAssuming axolotl is installed (if not, see our Installation Guide)\n\nDownload example configs:\n\naxolotl fetch examples\n\nRun the training:\n\naxolotl train examples/llama-3/lora-1b.yml\nThat’s it! Let’s understand what just happened.",
    "crumbs": [
      "Getting Started",
      "Quickstart"
    ]
  },
  {
    "objectID": "docs/getting-started.html#sec-understanding",
    "href": "docs/getting-started.html#sec-understanding",
    "title": "Quickstart",
    "section": "2 Understanding the Process",
    "text": "2 Understanding the Process\n\n2.1 The Configuration File\nThe YAML configuration file controls everything about your training. Here’s what (part of) our example config looks like:\nbase_model: NousResearch/Llama-3.2-1B\n\nload_in_8bit: true\nadapter: lora\n\ndatasets:\n  - path: teknium/GPT4-LLM-Cleaned\n    type: alpaca\ndataset_prepared_path: last_run_prepared\nval_set_size: 0.1\noutput_dir: ./outputs/lora-out\n\n\n\n\n\n\nTip\n\n\n\nload_in_8bit: true and adapter: lora enables LoRA adapter finetuning.\n\nTo perform Full finetuning, remove these two lines.\nTo perform QLoRA finetuning, replace with load_in_4bit: true and adapter: qlora.\n\n\n\nSee our Config options for more details.\n\n\n2.2 Training\nWhen you run axolotl train, Axolotl:\n\nDownloads the base model\n(If specified) applies QLoRA/LoRA adapter layers\nLoads and processes the dataset\nRuns the training loop\nSaves the trained model and / or LoRA weights",
    "crumbs": [
      "Getting Started",
      "Quickstart"
    ]
  },
  {
    "objectID": "docs/getting-started.html#sec-custom",
    "href": "docs/getting-started.html#sec-custom",
    "title": "Quickstart",
    "section": "3 Your First Custom Training",
    "text": "3 Your First Custom Training\nLet’s modify the example for your own data:\n\nCreate a new config file my_training.yml:\n\nbase_model: NousResearch/Nous-Hermes-llama-1b-v1\n\nload_in_8bit: true\nadapter: lora\n\n# Training settings\nmicro_batch_size: 2\nnum_epochs: 3\nlearning_rate: 0.0003\n\n# Your dataset\ndatasets:\n  - path: my_data.jsonl        # Your local data file\n    type: alpaca               # Or other format\nThis specific config is for LoRA fine-tuning a model with instruction tuning data using\nthe alpaca dataset format, which has the following format:\n{\n    \"instruction\": \"Write a description of alpacas.\",\n    \"input\": \"\",\n    \"output\": \"Alpacas are domesticated South American camelids...\"\n}\nPlease see our Dataset Formats for more dataset formats and how to\nformat them.\n\nPrepare your JSONL data in the specified format (in this case, the expected alpaca\nformat):\n\n{\"instruction\": \"Classify this text\", \"input\": \"I love this!\", \"output\": \"positive\"}\n{\"instruction\": \"Classify this text\", \"input\": \"Not good at all\", \"output\": \"negative\"}\n\nRun the training:\n\naxolotl train my_training.yml",
    "crumbs": [
      "Getting Started",
      "Quickstart"
    ]
  },
  {
    "objectID": "docs/getting-started.html#sec-common-tasks",
    "href": "docs/getting-started.html#sec-common-tasks",
    "title": "Quickstart",
    "section": "4 Common Tasks",
    "text": "4 Common Tasks\n\n\n\n\n\n\nTip\n\n\n\nThe same yaml file is used for training, inference, and merging.\n\n\n\n4.1 Testing Your Model\nAfter training, test your model:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\"\nMore details can be found in Inference.\n\n\n4.2 Using a UI\nLaunch a Gradio interface:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\" --gradio\n\n\n4.3 Preprocessing Data\nFor large datasets, preprocess first:\naxolotl preprocess my_training.yml\nPlease make sure to set dataset_prepared_path: in your config to set the path to save the prepared dataset.\nMore details can be found in Dataset Preprocessing.\n\n\n4.4 Merging LoRA weights\nTo merge the LoRA weights back into the base model, run:\naxolotl merge-lora my_training.yml --lora-model-dir=\"./outputs/lora-out\"\nThe merged model will be saved in the {output_dir}/merged directory.\nMore details can be found in Merging LoRA weights.",
    "crumbs": [
      "Getting Started",
      "Quickstart"
    ]
  },
  {
    "objectID": "docs/getting-started.html#sec-next-steps",
    "href": "docs/getting-started.html#sec-next-steps",
    "title": "Quickstart",
    "section": "5 Next Steps",
    "text": "5 Next Steps\nNow that you have the basics, you might want to:\n\nTry different model architectures\nExperiment with hyperparameters\nUse more advanced training methods\nScale up to larger models\n\nCheck our other guides for details on these topics:\n\nConfiguration Guide - Full configuration options\nDataset Loading - Loading datasets from various sources\nDataset Formats - Working with different data formats\nMulti-GPU Training\nMulti-Node Training",
    "crumbs": [
      "Getting Started",
      "Quickstart"
    ]
  },
  {
    "objectID": "docs/torchao.html",
    "href": "docs/torchao.html",
    "title": "PyTorch ao",
    "section": "",
    "text": "To use experimental optimizers (AdamWFp8, AdamW4bit, AdamW8bit) from Pytorch Ao, please install the package as shown below.\n\n\n\n\n\n\nTip\n\n\n\nSome experimental optimizers are already present in regular Pytorch, so please re-check if you actually need this package!\n\n\n\nInstallation\nStable Release from the PyTorch index\npip install torchao --extra-index-url https://download.pytorch.org/whl/cu121 # full options are cpu/cu118/cu121/cu124\nNightly release\npip install --pre torchao-nightly --index-url https://download.pytorch.org/whl/nightly/cu121 # full options are cpu/cu118/cu121/cu124",
    "crumbs": [
      "Advanced Features",
      "PyTorch ao"
    ]
  },
  {
    "objectID": "docs/dataset_preprocessing.html",
    "href": "docs/dataset_preprocessing.html",
    "title": "Dataset Preprocessing",
    "section": "",
    "text": "Dataset pre-processing is the step where Axolotl takes each dataset you’ve configured alongside\nthe dataset format and prompt strategies to:\n\nparse the dataset based on the dataset format\ntransform the dataset to how you would interact with the model based on the prompt strategy\ntokenize the dataset based on the configured model & tokenizer\nshuffle and merge multiple datasets together if using more than one\n\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling axolotl preprocess config.yaml --debug\nWhen training is started\n\n\n\nWhen training interactively or for sweeps\n(e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly\nslow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent\ntraining parameters so that it will intelligently pull from its cache when possible.\nThe path of the cache is controlled by dataset_prepared_path: and is often left blank in example\nYAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.\nIf dataset_prepared_path: is left empty, when training, the processed dataset will be cached in a\ndefault path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly\nsetting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed\ndata is in the cache.\n\n\n\nLet’s say you are writing a custom prompt strategy or using a user-defined\nprompt template. Because the trainer cannot readily detect these changes, we cannot change the\ncalculated hash value for the pre-processed dataset.\nIf you have dataset_prepared_path: ... set\nand change your prompt templating logic, it may not pick up the changes you made and you will be\ntraining over the old prompt.",
    "crumbs": [
      "Core Concepts",
      "Dataset Preprocessing"
    ]
  },
  {
    "objectID": "docs/dataset_preprocessing.html#overview",
    "href": "docs/dataset_preprocessing.html#overview",
    "title": "Dataset Preprocessing",
    "section": "",
    "text": "Dataset pre-processing is the step where Axolotl takes each dataset you’ve configured alongside\nthe dataset format and prompt strategies to:\n\nparse the dataset based on the dataset format\ntransform the dataset to how you would interact with the model based on the prompt strategy\ntokenize the dataset based on the configured model & tokenizer\nshuffle and merge multiple datasets together if using more than one\n\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling axolotl preprocess config.yaml --debug\nWhen training is started\n\n\n\nWhen training interactively or for sweeps\n(e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly\nslow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent\ntraining parameters so that it will intelligently pull from its cache when possible.\nThe path of the cache is controlled by dataset_prepared_path: and is often left blank in example\nYAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.\nIf dataset_prepared_path: is left empty, when training, the processed dataset will be cached in a\ndefault path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly\nsetting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed\ndata is in the cache.\n\n\n\nLet’s say you are writing a custom prompt strategy or using a user-defined\nprompt template. Because the trainer cannot readily detect these changes, we cannot change the\ncalculated hash value for the pre-processed dataset.\nIf you have dataset_prepared_path: ... set\nand change your prompt templating logic, it may not pick up the changes you made and you will be\ntraining over the old prompt.",
    "crumbs": [
      "Core Concepts",
      "Dataset Preprocessing"
    ]
  },
  {
    "objectID": "docs/docker.html",
    "href": "docs/docker.html",
    "title": "Docker",
    "section": "",
    "text": "This section describes the different Docker images that are released by AxolotlAI at Docker Hub.",
    "crumbs": [
      "Deployments",
      "Docker"
    ]
  },
  {
    "objectID": "docs/docker.html#base",
    "href": "docs/docker.html#base",
    "title": "Docker",
    "section": "Base",
    "text": "Base\nThe base image is the most minimal image that can install Axolotl. It is based on the nvidia/cuda image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.\n\nImage\naxolotlai/axolotl-base\nLink: Docker Hub\n\n\nTags format\nmain-base-py{python_version}-cu{cuda_version}-{pytorch_version}\nTags examples:\n\nmain-base-py3.11-cu128-2.7.0\nmain-base-py3.11-cu126-2.7.0\nmain-base-py3.11-cu124-2.6.0\nmain-base-py3.11-cu124-2.5.1",
    "crumbs": [
      "Deployments",
      "Docker"
    ]
  },
  {
    "objectID": "docs/docker.html#main",
    "href": "docs/docker.html#main",
    "title": "Docker",
    "section": "Main",
    "text": "Main\nThe main image is the image that is used to run Axolotl. It is based on the axolotlai/axolotl-base image and includes the Axolotl codebase, dependencies, and more.\n\nImage\naxolotlai/axolotl\nLink: Docker Hub\n\n\nTags format\n# on push to main\nmain-py{python_version}-cu{cuda_version}-{pytorch_version}\n\n# latest main (currently torch 2.6.0, python 3.11, cuda 12.4)\nmain-latest\n\n# nightly build\n{branch}-{date_in_YYYYMMDD}-py{python_version}-cu{cuda_version}-{pytorch_version}\n\n# tagged release\n{version}\n\n\n\n\n\n\nTip\n\n\n\nThere may be some extra tags appended to the image, like -vllm which installs those packages.\n\n\nTags examples:\n\nmain-py3.11-cu126-2.7.0\nmain-py3.11-cu124-2.6.0\nmain-py3.11-cu124-2.5.1\nmain-latest\nmain-20250303-py3.11-cu124-2.6.0\nmain-20250303-py3.11-cu124-2.5.1\n0.9.2",
    "crumbs": [
      "Deployments",
      "Docker"
    ]
  },
  {
    "objectID": "docs/docker.html#cloud",
    "href": "docs/docker.html#cloud",
    "title": "Docker",
    "section": "Cloud",
    "text": "Cloud\nThe cloud image is the image that is used to run Axolotl in the cloud. It is based on the axolotlai/axolotl image and sets ENV variables like HuggingFace cache directories for volume mounts, tmux, and more for different cloud providers.\n\n\n\n\n\n\nTip\n\n\n\nJupyter lab is run by default. Set JUPYTER_DISABLE=1 in the environment variables to disable it.\n\n\n\nImage\naxolotlai/axolotl-cloud\nLink: Docker Hub\n\n\nTags format\nThis uses the same tags as the main image.\n\n\nEnvironment variables\n\nJUPYTER_DISABLE: Disable Jupyter lab.\nJUPYTER_PASSWORD: Set a password for the Jupyter lab.\nPUBLIC_KEY / SSH_KEY: Add a public key for the SSH service.\n\n\n\nVolume mounts\n\n\n\n\n\n\nTip\n\n\n\nWe recommend mounting volumes to /workspace/data for data persistence. /workspace/axolotl contains the source code and is ephemeral.\n\n\n\n/workspace/data/axolotl-artifacts: Directory to store Axolotl artifacts.\n/workspace/data/huggingface-cache: Directory to store HuggingFace cache.",
    "crumbs": [
      "Deployments",
      "Docker"
    ]
  },
  {
    "objectID": "docs/docker.html#cloud-no-tmux",
    "href": "docs/docker.html#cloud-no-tmux",
    "title": "Docker",
    "section": "Cloud-no-tmux",
    "text": "Cloud-no-tmux\nThis is the same as the cloud image but without tmux.\n\nImage\naxolotlai/axolotl-cloud-term\nLink: Docker Hub\n\n\n\n\n\n\nNote\n\n\n\nThe naming may be a bit confusing as it has -term appended to the end.\n\n\n\n\nTags format\nThis uses the same tags as the cloud image.",
    "crumbs": [
      "Deployments",
      "Docker"
    ]
  },
  {
    "objectID": "docs/ray-integration.html",
    "href": "docs/ray-integration.html",
    "title": "Ray Train",
    "section": "",
    "text": "Axolotl supports using Ray as an alternative to accelerate for orchestrating training. This is especially useful for multi-node training since you only have to setup code and dependencies in a single node and launch training as if you were using a single node.\nWith the --use-ray CLI flag, Axolotl will use Ray Train’s TorchTrainer to run training.",
    "crumbs": [
      "Deployments",
      "Ray Train"
    ]
  },
  {
    "objectID": "docs/ray-integration.html#ray-cluster-setup",
    "href": "docs/ray-integration.html#ray-cluster-setup",
    "title": "Ray Train",
    "section": "Ray cluster setup",
    "text": "Ray cluster setup\nA prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs here.\nEvery Ray cluster has one head node and a set of worker nodes. The head node is just like any other worker node, but it also runs certain special processes related to scheduling and orchestration. Ray-enabled scripts are run on the head node and depending on the resources (number of CPUs, GPUs, etc) they request, will be scheduled to run certain tasks on the worker nodes. For more on key concepts behind a Ray cluster, you can refer this doc.",
    "crumbs": [
      "Deployments",
      "Ray Train"
    ]
  },
  {
    "objectID": "docs/ray-integration.html#sanity-check",
    "href": "docs/ray-integration.html#sanity-check",
    "title": "Ray Train",
    "section": "Sanity check",
    "text": "Sanity check\nTo run a sanity check on whether your ray cluster is setup properly, execute the following on the head node:\nray status\nThe output should have a summary of your Ray cluster - list of all the nodes in your cluster, the number of CPUs and GPUs in your cluster, etc. For example, if you have a cluster with 1 CPU-only head node and 2 4xL40S worker nodes, the output can look like this:\nNode status\n---------------------------------------------------------------\nActive:\n 1 head\nIdle:\n 2 4xL40S:48CPU-384GB\nPending:\n (no pending nodes)\nRecent failures:\n (no failures)\n\nResources\n---------------------------------------------------------------\nUsage:\n 0.0/96.0 CPU\n 0.0/8.0 GPU\n 0B/800.00GiB memory\n 0B/229.57GiB object_store_memory\n\nDemands:\n (no resource demands)\nYou should also be able to see the same on the Ray dashboard.",
    "crumbs": [
      "Deployments",
      "Ray Train"
    ]
  },
  {
    "objectID": "docs/ray-integration.html#configuring-training-with-ray-train",
    "href": "docs/ray-integration.html#configuring-training-with-ray-train",
    "title": "Ray Train",
    "section": "Configuring training with Ray Train",
    "text": "Configuring training with Ray Train\nYou can find an example configuration at configs/llama-3/lora-1b-ray.yaml.\nThe key parameters to note here are:\nuse_ray: true\nray_num_workers: 4\n# optional\nresources_per_worker:\n    GPU: 1\n\nuse_ray: This is the flag that enables the Ray Train integration. You can either use the corresponding --use-ray flag in the CLI or set use_ray in the config file.\nray_num_workers: This is the number of workers/GPUs to use for training.\nresources_per_worker: This is the Ray resource request for each worker. This can be used to request a specific GPU type or a custom resource for each worker. For example, if your ray cluster has GPUs of different types, and you only want to use NVIDIA L40S GPUs, you can do\n\nresources_per_worker:\n    accelerator_type:L40S: 0.001",
    "crumbs": [
      "Deployments",
      "Ray Train"
    ]
  },
  {
    "objectID": "docs/ray-integration.html#launching-training",
    "href": "docs/ray-integration.html#launching-training",
    "title": "Ray Train",
    "section": "Launching training",
    "text": "Launching training\nYou can simply run the following command on the head node:\naxolotl train examples/llama-3/lora-1b-ray.yml --use-ray\nThis will launch training on the head node and workers will be scheduled automatically by Ray Train to run on the appropriate head or worker nodes.\nYou can also monitor training progress on the Ray dashboard.\nComing back to the example on a Ray cluster with 1 head node and 2 4xL40S worker nodes, let’s say you want to make use of all 8 GPUs. You would be able to just set ray_num_workers: 8 and run the previous command. The Cluster tab will show the following:\n\n\n\nRay dashboard",
    "crumbs": [
      "Deployments",
      "Ray Train"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html",
    "href": "docs/dataset-formats/inst_tune.html",
    "title": "Instruction Tuning",
    "section": "",
    "text": "instruction; input(optional)\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"input\": \"...\", \"output\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#alpaca",
    "href": "docs/dataset-formats/inst_tune.html#alpaca",
    "title": "Instruction Tuning",
    "section": "",
    "text": "instruction; input(optional)\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"input\": \"...\", \"output\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#jeopardy",
    "href": "docs/dataset-formats/inst_tune.html#jeopardy",
    "title": "Instruction Tuning",
    "section": "jeopardy",
    "text": "jeopardy\nquestion and answer\n\n\ndata.jsonl\n\n{\"question\": \"...\", \"category\": \"...\", \"answer\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#oasst",
    "href": "docs/dataset-formats/inst_tune.html#oasst",
    "title": "Instruction Tuning",
    "section": "oasst",
    "text": "oasst\ninstruction\n\n\ndata.jsonl\n\n{\"INSTRUCTION\": \"...\", \"RESPONSE\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#gpteacher",
    "href": "docs/dataset-formats/inst_tune.html#gpteacher",
    "title": "Instruction Tuning",
    "section": "gpteacher",
    "text": "gpteacher\ninstruction; input(optional)\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"input\": \"...\", \"response\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#reflection",
    "href": "docs/dataset-formats/inst_tune.html#reflection",
    "title": "Instruction Tuning",
    "section": "reflection",
    "text": "reflection\ninstruction with reflect; input(optional)\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"input\": \"...\", \"output\": \"...\", \"reflection\": \"...\", \"corrected\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#explainchoice",
    "href": "docs/dataset-formats/inst_tune.html#explainchoice",
    "title": "Instruction Tuning",
    "section": "explainchoice",
    "text": "explainchoice\nquestion, choices, (solution OR explanation)\n\n\ndata.jsonl\n\n{\"question\": \"...\", \"choices\": [\"...\"], \"solution\": \"...\", \"explanation\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#concisechoice",
    "href": "docs/dataset-formats/inst_tune.html#concisechoice",
    "title": "Instruction Tuning",
    "section": "concisechoice",
    "text": "concisechoice\nquestion, choices, (solution OR explanation)\n\n\ndata.jsonl\n\n{\"question\": \"...\", \"choices\": [\"...\"], \"solution\": \"...\", \"explanation\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#summarizetldr",
    "href": "docs/dataset-formats/inst_tune.html#summarizetldr",
    "title": "Instruction Tuning",
    "section": "summarizetldr",
    "text": "summarizetldr\narticle and summary\n\n\ndata.jsonl\n\n{\"article\": \"...\", \"summary\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#alpaca_chat",
    "href": "docs/dataset-formats/inst_tune.html#alpaca_chat",
    "title": "Instruction Tuning",
    "section": "alpaca_chat",
    "text": "alpaca_chat\nbasic instruct for alpaca chat\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"input\": \"...\", \"response\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#alpaca_chat.load_qa",
    "href": "docs/dataset-formats/inst_tune.html#alpaca_chat.load_qa",
    "title": "Instruction Tuning",
    "section": "alpaca_chat.load_qa",
    "text": "alpaca_chat.load_qa\nquestion and answer for alpaca chat\n\n\ndata.jsonl\n\n{\"question\": \"...\", \"answer\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#alpaca_chat.load_concise",
    "href": "docs/dataset-formats/inst_tune.html#alpaca_chat.load_concise",
    "title": "Instruction Tuning",
    "section": "alpaca_chat.load_concise",
    "text": "alpaca_chat.load_concise\nquestion and answer for alpaca chat, for concise answers\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"input\": \"...\", \"response\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#alpaca_chat.load_camel_ai",
    "href": "docs/dataset-formats/inst_tune.html#alpaca_chat.load_camel_ai",
    "title": "Instruction Tuning",
    "section": "alpaca_chat.load_camel_ai",
    "text": "alpaca_chat.load_camel_ai\nquestion and answer for alpaca chat, for load_camel_ai\n\n\ndata.jsonl\n\n{\"message_1\": \"...\", \"message_2\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#alpaca_w_system.load_open_orca",
    "href": "docs/dataset-formats/inst_tune.html#alpaca_w_system.load_open_orca",
    "title": "Instruction Tuning",
    "section": "alpaca_w_system.load_open_orca",
    "text": "alpaca_w_system.load_open_orca\nsupport for open orca datasets with included system prompts, instruct\n\n\ndata.jsonl\n\n{\"system_prompt\": \"...\", \"question\": \"...\", \"response\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#context_qa",
    "href": "docs/dataset-formats/inst_tune.html#context_qa",
    "title": "Instruction Tuning",
    "section": "context_qa",
    "text": "context_qa\nin context question answering from an article\n\n\ndata.jsonl\n\n{\"article\": \"...\", \"question\": \"...\", \"answer\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#context_qa.load_v2",
    "href": "docs/dataset-formats/inst_tune.html#context_qa.load_v2",
    "title": "Instruction Tuning",
    "section": "context_qa.load_v2",
    "text": "context_qa.load_v2\nin context question answering (alternate)\n\n\ndata.jsonl\n\n{\"context\": \"...\", \"question\": \"...\", \"answer\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#context_qa.load_404",
    "href": "docs/dataset-formats/inst_tune.html#context_qa.load_404",
    "title": "Instruction Tuning",
    "section": "context_qa.load_404",
    "text": "context_qa.load_404\nin context question answering from an article, with default response for no answer from context\n\n\ndata.jsonl\n\n{\"article\": \"...\", \"unanswerable_question\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#creative_acr.load_answer",
    "href": "docs/dataset-formats/inst_tune.html#creative_acr.load_answer",
    "title": "Instruction Tuning",
    "section": "creative_acr.load_answer",
    "text": "creative_acr.load_answer\ninstruction and revision\n\n\ndata.jsonl\n\n{\"instruction\": \"...\", \"revision\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#creative_acr.load_critique",
    "href": "docs/dataset-formats/inst_tune.html#creative_acr.load_critique",
    "title": "Instruction Tuning",
    "section": "creative_acr.load_critique",
    "text": "creative_acr.load_critique\ncritique\n\n\ndata.jsonl\n\n{\"scores\": \"...\", \"critiques\": \"...\", \"instruction\": \"...\", \"answer\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#creative_acr.load_revise",
    "href": "docs/dataset-formats/inst_tune.html#creative_acr.load_revise",
    "title": "Instruction Tuning",
    "section": "creative_acr.load_revise",
    "text": "creative_acr.load_revise\ncritique and revise\n\n\ndata.jsonl\n\n{\"scores\": \"...\", \"critiques\": \"...\", \"instruction\": \"...\", \"answer\": \"...\", \"revision\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#metharme",
    "href": "docs/dataset-formats/inst_tune.html#metharme",
    "title": "Instruction Tuning",
    "section": "metharme",
    "text": "metharme\ninstruction, adds additional eos tokens\n\n\ndata.jsonl\n\n{\"prompt\": \"...\", \"generation\": \"...\"}",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/inst_tune.html#how-to-add-custom-prompt-format",
    "href": "docs/dataset-formats/inst_tune.html#how-to-add-custom-prompt-format",
    "title": "Instruction Tuning",
    "section": "How to add custom prompt format",
    "text": "How to add custom prompt format\nFor a dataset that is preprocessed for instruction purposes:\n\n\ndata.jsonl\n\n{\"input\": \"...\", \"output\": \"...\"}\n\nYou can use this example in your YAML config:\n\n\nconfig.yaml\n\ndatasets:\n  - path: repo\n    type:\n      system_prompt: \"\"\n      field_system: system\n      field_instruction: input\n      field_output: output\n      format: \"[INST] {instruction} [/INST]\"\n      no_input_format: \"[INST] {instruction} [/INST]\"\n\nSee full config options under here.",
    "crumbs": [
      "Dataset Formats",
      "Instruction Tuning"
    ]
  },
  {
    "objectID": "docs/dataset-formats/stepwise_supervised.html",
    "href": "docs/dataset-formats/stepwise_supervised.html",
    "title": "Stepwise Supervised Format",
    "section": "",
    "text": "The stepwise supervised format is designed for chain-of-thought (COT) reasoning\ndatasets where each example contains multiple completion steps and a preference label\nfor each step.\n\n\nHere’s a simple example of a stepwise supervised dataset entry:\n{\n  \"prompt\": \"Which number is larger, 9.8 or 9.11?\",\n  \"completions\": [\n    \"The fractional part of 9.8 is 0.8, while the fractional part of 9.11 is 0.11.\",\n    \"Since 0.11 is greater than 0.8, the number 9.11 is larger than 9.8.\"\n  ],\n  \"labels\": [true, false]\n}",
    "crumbs": [
      "Dataset Formats",
      "Stepwise Supervised Format"
    ]
  },
  {
    "objectID": "docs/dataset-formats/stepwise_supervised.html#stepwise-supervised",
    "href": "docs/dataset-formats/stepwise_supervised.html#stepwise-supervised",
    "title": "Stepwise Supervised Format",
    "section": "",
    "text": "The stepwise supervised format is designed for chain-of-thought (COT) reasoning\ndatasets where each example contains multiple completion steps and a preference label\nfor each step.\n\n\nHere’s a simple example of a stepwise supervised dataset entry:\n{\n  \"prompt\": \"Which number is larger, 9.8 or 9.11?\",\n  \"completions\": [\n    \"The fractional part of 9.8 is 0.8, while the fractional part of 9.11 is 0.11.\",\n    \"Since 0.11 is greater than 0.8, the number 9.11 is larger than 9.8.\"\n  ],\n  \"labels\": [true, false]\n}",
    "crumbs": [
      "Dataset Formats",
      "Stepwise Supervised Format"
    ]
  },
  {
    "objectID": "docs/dataset-formats/index.html",
    "href": "docs/dataset-formats/index.html",
    "title": "Dataset Formats",
    "section": "",
    "text": "Axolotl is a training framework that aims to make the process convenient yet flexible to users by simply passing a config yaml file.\nAs there are a lot of available options in Axolotl, this guide aims to provide an simplify the user experience to choosing the proper choice.\nAxolotl supports 3 kinds of training methods: pre-training, supervised fine-tuning, and preference-based post-training (e.g. DPO, ORPO, PRMs). Each method has their own dataset format which are described below.",
    "crumbs": [
      "Dataset Formats"
    ]
  },
  {
    "objectID": "docs/dataset-formats/index.html#pre-training",
    "href": "docs/dataset-formats/index.html#pre-training",
    "title": "Dataset Formats",
    "section": "Pre-training",
    "text": "Pre-training\nWhen aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports streaming to only load batches into memory at a time.\nA sample format for a pre-training dataset is as follows:\n{\"text\": \"first row\"}\n{\"text\": \"second row\"}\n...\nIt is typically recommended to save your dataset as .jsonl due to its flexibility and simplicity.\nAxolotl supports loading from a Hugging Face hub repo or from local files.\n\nPre-training from Hugging Face hub datasets\nAs an example, to train using a Hugging Face dataset hf_org/name, you can pass the following config:\npretraining_dataset: hf_org/name\n\n\nPre-training from local dataset files\nGiven a few corpus files: A.jsonl, B.jsonl, and C.jsonl, your config will look like the below:\npretraining_dataset:\n  - path: json\n    data_files:\n      - A.jsonl\n      - B.jsonl\n      - C.jsonl\nWhile we recommend .jsonl, you can also use the other formats (csv, parquet, arrow, SQL, Webdataset) that are supported by Dataset.load_dataset\n\n\nPre-training without streaming\nOn the rare case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the completion format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.\nOne benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.\nFrom Hugging Face:\ndatasets:\n  - path: hf_org/name\n    type: completion\nFrom local files:\ndatasets:\n  - path: A.jsonl\n    type: completion\n\n  - path: B.jsonl\n    type: completion\n\n\n\n\n\n\nImportant\n\n\n\nFor completion only, Axolotl would split texts if it exceeds the context length into multiple smaller prompts. If you are interested in having this for pretraining_dataset too, please let us know or help make a PR!\n\n\n\n\nPre-training dataset configuration tips\n\nSetting max_steps\nWhen using streaming for large datasets, Axolotl does not know in advance how large the dataset is and does not know when to stop.\nTherefore, it is necessary to set max_steps: int in your config for pre-training to run, so that Axolotl knows when to stop training.\nOne step is equal to sequence_len * micro_batch_size * gradient_accumulation_steps * total_num_gpus tokens.\n\n\nGroup_by_length\nIt is recommended to leave this off if downloading from Hugging Face hub as it would download the entire dataset which can be very large.\n\n\n\nReference\nPlease see docs here.",
    "crumbs": [
      "Dataset Formats"
    ]
  },
  {
    "objectID": "docs/dataset-formats/index.html#supervised-fine-tuning-sft",
    "href": "docs/dataset-formats/index.html#supervised-fine-tuning-sft",
    "title": "Dataset Formats",
    "section": "Supervised fine-tuning (SFT)",
    "text": "Supervised fine-tuning (SFT)\nSupervised fine-tuning is the process of training models to respond to an instruction or chat input.\nAs there are a wide variety of dataset formats, Axolotl tries to support a majority of the formats available in public datasets.\nAxolotl provides four approaches for loading datasets, however, it’s easier to work backwards from the dataset you have available to figure out which approach to use.\nA flow chart is as follows:\n\nDo you already have the dataset tokenized? If yes, check Pre-Tokenized Dataset.\nDo you want to format the dataset yourself and manually choose each section to mask? If yes, check Template Free Dataset\nIs your dataset in a “conversation” format, containing a list[messages]? If yes, check Conversation Dataset\nIs your dataset in an “instruct” format, containing { instruction, response }? If yes, check Instruction Dataset\n\nIf you went through the flow chart and did not find one that matches, it is recommended to preprocess your dataset into one of the above or create a thread on Github Discussion.\n\n\n\n\n\n\nTip\n\n\n\nYou can mix and match within each approach or across approaches to train a model on a variety of datasets.\n\n\n\nPre-Tokenized Dataset\nWe suggest this approach when you want to bring your own tokenized dataset.\nAxolotl expects the dataset to have three keys:\n\ninput_ids: from tokenizing formatted prompt\nattention_mask: for masking padding. If you don’t add padding, it would be equal to len(input_ids) * [1]\nlabels: this is the same as input_ids, however, if you want to mask certain tokens, you would set those indices to -100.\n\n\n\n\n\n\n\nTip\n\n\n\nMake sure to add BOS/EOS tokens to your prompt and mask it appropriately.\n\n\nA config for this would look like:\ndatasets:\n  - path: A.jsonl\n    type:\n\n\n\n\n\n\nNote\n\n\n\ntype: is empty!\n\n\nReference: Pre-Tokenized Dataset Documentation.\n\n\nTemplate Free Dataset\nWe reccomend this approach when you want granular control over the prompt formatting, special tokens, and masking, whilst letting Axolotl handle the tokenization. This is very useful if your dataset has unique prompts that differ across samples and where one single general template wouldn’t suffice.\nIn the example below, you could see that there is no proper structure. At the same time, it’s very flexible as there are no constraints on how your prompt can look.\n{\n    \"segments\": [\n        {\n            \"label\": true,\n            \"text\": \"<s>Hello\\n\"\n        },\n        {\n            \"label\": true,\n            \"text\": \"hi there!. \"\n        },\n        {\n            \"label\": false,\n            \"text\": \"goodbye \"\n        },\n        {\n            \"label\": true,\n            \"text\": \"farewell</s>\"\n        }\n    ]\n}\nEach prompt must be have a key called segments which is a list of { text, label }.\ndatasets:\n  - path: A.jsonl\n    type: input_output\nReference: Template Free Documentation.\n\n\nConversation Dataset\nconversation messages are a list of messages which usually contain a role and content key.\n\n\n\n\n\n\nTip\n\n\n\nFun fact: Axolotl synonymously refers to “chat” messages as conversation messages due to how FastChat initially used this term to build a widely used fastchat conversation method for formatting chat messages prior to the creation of chat_templates.\n\n\n\nWhat are chat_templates?\nThe current most popular and convenient method for inference is to use chat_templates for formatting prompts. Axolotl supports using chat_templates for training to ensure that the model performs in the same environment as in inference.\nHere’s a quick rundown on chat_template: A chat_template is a Jinja2 template which formats a list of messages into a prompt.\nAn example of a prompt formatted into a popular template called ChatML can be seen below:\nSingle prompt (pretty-printed):\n{\n    \"messages\": [\n        {\n            \"role\": \"user\",\n            \"content\": \"Hi\"\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": \"How can I help you?\"\n        },\n        {\n            \"role\": \"user\",\n            \"content\": \"Can you add 3+5?\"\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": \"The answer is 8.\"\n        }\n    ]\n}\nThe ChatML template is as follows:\n{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}\nThe above prompt formatted into this template will result in:\n<|im_start|>user\nHi<|im_end|>\n<|im_start|>assistant\nHow can I help you?<|im_end|>\n<|im_start|>user\nCan you add 3+5?<|im_end|>\n<|im_start|>assistant\nThe answer is 8.<|im_end|>\nBy using delimiters (<|im_start|> and <|im_end|>), a prompt separates different speakers which helps the model identify which portion belongs to whom.\n\n\nCommon Conversation Dataset formats\nOlder conversation datasets with the following format are colloquially called sharegpt datasets.\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\nNewer conversation datasets usually follow the OpenAI format.\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}]}\nAxolotl supports both as well as allowing customization of any kind of key.\n\n\nChat Template Usage\nTo properly use this method, it is important to identify three things:\n\nWhich chat_template would you use?\nWhat are the keys in your dataset, and what are the possible roles? For example, in OpenAI format, the keys would be messages, role, and content, respectively, whereas the possible roles are system, user, and assistant.\nWhat do you want to mask? For instance, only assistant messages, only last message, or nothing.\n\n\nChoosing a chat_template\nThere are a lot of chat_templates out there. Axolotl supports the common ones: supported chat templates. For example, to use ChatML, it would be chat_template: chatml.\nHowever, it is also possible to use the already configured template within the tokenizer by specifying chat_template: tokenizer_default. If you want a fallback (in case some tokenizer does not have it pre-configured), you can do chat_template: tokenizer_default_fallback_chatml to fallback to the ChatML template if a tokenizer template was not found.\nOne last but powerful approach is to bring your own template. This can be set via:\nchat_template_jinja: # your template\n\n\nSetting chat_template dataset keys\nWe currently default to OpenAI format for dataset keys, so if that’s your current dataset format, there’s nothing to do here.\nIf your dataset format is different, here are the keys you should check (with their defaults):\ndatasets:\n    ...\n    field_messages: messages  # this should point to the key containing the list of conversations\n    message_property_mappings:  # this is a mapping from keys in your dataset to keys in chat_template\n      role: role\n      content: content\nIn some chat_templates (e.g. Gemma), the roles are hardcoded to user and assistant. Consequently, you may find it necessary to map the roles in your dataset to these above. We currently have some defaults that should work for common datasets, but if you get a KeyError, it would be necessary to add mapping for your roles. Here is an example of how it would look like:\ndatasets:\n    ...\n    roles:\n      assistant:\n        - gpt\n        - model\n      user:\n        - human\nIn the example above, all gpt and model values are converted to assistant. All human values are converted to user.\n\n\nHandling masking\nThe common use case for chat_template is for chat messages, therefore, it is common to mask all non-assistant messages. Assistant messages refer to the bot messages that you want the model to learn on.\nTo train on all assistant messages, you would set the following configs.\ndatasets:\n    ...\n    roles_to_train: [\"assistant\"]\n    train_on_eos: \"turn\"\nThe train_on_eos config means that it would mask all EOS tokens for turns that aren’t assistant-turns. The other options are: all and last to choose which EOS to train on.\nPerhaps, you want to train on assistant and narrator roles, you can simply add narrator to the list of roles_to_train. You would also need to add it to the mapping of roles above.\ndatasets:\n    ...\n    roles_to_train: [\"assistant\", \"narrator\"]\n    roles:\n      assistant:\n        - gpt\n        - model\n      user:\n        - human\n      narrator: [\"narrator\"]\n\n\n\n\n\n\nTip\n\n\n\nAs chat_templates may use hardcoded EOS/EOT tokens that are different from the tokenizer’s EOS, it is highly recommended to set them. For example, ChatML uses <|im_end|> to end turns.\nspecial_tokens:\n  eos_token: <|im_end|>\n\n\n\n\nApplying chat_template\nOnce all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your custom dataset.\ndatasets:\n  - path: A.jsonl\n    type: chat_template\n\n    # step 1\n    chat_template: chatml\n\n    # step 2\n    field_messages: messages\n    message_property_mappings:\n      role: role\n      content: content\n\n    roles:\n      assistant:\n        - gpt\n        - model\n        - assistant\n      user:\n        - human\n        - user\n\n    # step 3\n    roles_to_train: [\"assistant\"]\n    train_on_eos: \"turn\"\n\nspecial_tokens:\n  eos_token: <|im_end|>\nIf this config were to be applied to the sample dataset above, the output would look as such (which can be retrieved via axolotl preprocess config.yaml --debug):\n<|im_start|>(-100, 128256) user(-100, 882)\n(-100, 198) Hi(-100, 13347) <|im_end|>(-100, 128257)\n(-100, 198) <|im_start|>(-100, 128256) assistant(-100, 78191)\n(-100, 198) How(4438, 4438)  can(649, 649)  I(358, 358)  help(1520, 1520)  you(499, 499) ?(30, 30) <|im_end|>(128257, 128257)\n(-100, 198) <|im_start|>(-100, 128256) user(-100, 882)\n(-100, 198) Can(-100, 6854)  you(-100, 499)  add(-100, 923)  (-100, 220) 3(-100, 18) +(-100, 10) 5(-100, 20) ?(-100, 30) <|im_end|>(-100, 128257)\n(-100, 198) <|im_start|>(-100, 128256) assistant(-100, 78191)\n(-100, 198) The(791, 791)  answer(4320, 4320)  is(374, 374)  (220, 220) 8(23, 23) .(13, 13) <|im_end|>(128257, 128257)\n(-100, 198)\nThe first number refers to the label, the second refers to the token_id. For example, -100 labels appear on non-assistant portions, meaning that they are masked during. For assistant portions, the label is the same as the token_id.\n\n\n\n\n\n\nNote\n\n\n\nIf during preprocess, there are a lot of warnings of Could not find content __ boundary, please check the FAQ section for chat_templates.\n\n\n\n\n\nReference\nPlease see docs here.\n\n\n\nInstruction Dataset\nInstruction datasets are used to train instruction-following models and comprise a prompt, containing an instruction, and a single response. In contrast to chat datasets which may be multi-turn, instruct datasets are typically single-turn.\nAn example is of a common format called Alpaca:\n{\"instruction\": \"...\", \"input\": \"...\", \"output\": \"...\"}\nUsing those keys, a prompt can be built based on it.\nBelow is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n{output}\nThis can be configured as such:\ndatasets:\n  - path: A.jsonl\n    type: alpaca\nAxolotl supports many kinds of instruction dataset. All of them can be found in the Instruction Dataset Documentation with their respective type and sample row format.\n\nCustom Instruct Prompt Format\nDue to the myriad possibilities of instruction formats, Axolotl allows customizing your own instruction format without having to dive into the code directly.\nIn the example below, a sample row is used to output in mistral_v1 format.\n{\"input\": \"...\", \"output\": \"...\"}\ndatasets:\n  - path: repo\n    type:\n      system_prompt: \"\"\n\n      field_system:\n      field_instruction: input\n      field_input:\n      field_output: output\n\n      # multi-line example with input\n      format: |-\n        [INST] {instruction} {input} [/INST]\n\n      # single-line example without input\n      no_input_format: \"[INST] {instruction} [/INST]\"\nThe config sets that the field_instruction is actually named input, and the field_input is empty as we don’t have an input in this sample. Generally, instruction can be thought as the question to the model, and input as the additional information with output being the response. It is not necessary to have an input nor system. In the end, the most important part is to understand what format you want it to look like and how you can customize this to your use case.\nReference: Custom Instruct Prompt Format Documentation.",
    "crumbs": [
      "Dataset Formats"
    ]
  },
  {
    "objectID": "docs/dataset-formats/index.html#reinforcement-learning-from-human-feedback-rlhf",
    "href": "docs/dataset-formats/index.html#reinforcement-learning-from-human-feedback-rlhf",
    "title": "Dataset Formats",
    "section": "Reinforcement Learning from Human Feedback (RLHF)",
    "text": "Reinforcement Learning from Human Feedback (RLHF)\nAs there are multiple RLHF methods with their own dataset requirements. Please see RLHF documentation for more detail.",
    "crumbs": [
      "Dataset Formats"
    ]
  },
  {
    "objectID": "docs/multimodal.html",
    "href": "docs/multimodal.html",
    "title": "MultiModal / Vision Language Models (BETA)",
    "section": "",
    "text": "Mllama\nLlama4\nPixtral\nLlava-1.5\nMistral-Small-3.1\nGemma-3\nQwen2-VL\nQwen2.5-VL",
    "crumbs": [
      "How To Guides",
      "MultiModal / Vision Language Models (BETA)"
    ]
  },
  {
    "objectID": "docs/multimodal.html#supported-models",
    "href": "docs/multimodal.html#supported-models",
    "title": "MultiModal / Vision Language Models (BETA)",
    "section": "",
    "text": "Mllama\nLlama4\nPixtral\nLlava-1.5\nMistral-Small-3.1\nGemma-3\nQwen2-VL\nQwen2.5-VL",
    "crumbs": [
      "How To Guides",
      "MultiModal / Vision Language Models (BETA)"
    ]
  },
  {
    "objectID": "docs/multimodal.html#usage",
    "href": "docs/multimodal.html#usage",
    "title": "MultiModal / Vision Language Models (BETA)",
    "section": "Usage",
    "text": "Usage\nMultimodal support is limited and doesn’t have full feature parity.\nHere are the hyperparams you’ll need to use to finetune a multimodal model.\nprocessor_type: AutoProcessor\n\nskip_prepare_dataset: true\nremove_unused_columns: false  # leave columns in place as they are needed to handle image embeddings during training\nsample_packing: false  # not yet supported with multimodal\n\nchat_template:  # see in next section\n\n# example dataset\ndatasets:\n  - path: HuggingFaceH4/llava-instruct-mix-vsft\n    type: chat_template\n    split: train[:1%]\n    field_messages: messages\n\n# (optional) if doing lora, only finetune the Language model,\n# leave the vision model and vision tower frozen\n# load_in_8bit: true\nadapter: lora\nlora_target_modules: 'language_model.model.layers.[\\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'\n\n# (optional) if you want to resize images to a set size\nimage_size: 512\nimage_resize_algorithm: bilinear\nPlease see examples folder for full configs.\n\n\n\n\n\n\nWarning\n\n\n\nSome of our chat_templates have been extended to support broader dataset types. This should not break any existing configs.\n\n\n\nMllama\nbase_model: meta-llama/Llama-3.2-11B-Vision-Instruct\n\nchat_template: llama3_2_vision\n\n\nLlama4\nbase_model: meta-llama/Llama-4-Scout-17B-16E-Instruct\n\nchat_template: llama4\n\n\nPixtral\nbase_model: mistralai/Pixtral-12B-2409\n\nchat_template: pixtral\n\n\nLlava-1.5\nbase_model: llava-hf/llava-1.5-7b-hf\n\nchat_template: llava\n\n\nMistral-Small-3.1\nbase_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503\n\nchat_template: mistral_v7_tekken\n\n\nGemma-3\n\n\n\n\n\n\nTip\n\n\n\nThe Gemma3-1B model is a text-only model, so please train as regular text model.\n\n\nFor multi-modal 4B/12B/27B models, use the following config:\nbase_model: google/gemma-3-4b-it\n\nchat_template: gemma3\n\n\nQwen2-VL\nbase_model: Qwen/Qwen2-VL-7B-Instruct\n\nchat_template: qwen2_vl\n\n\nQwen2.5-VL\nbase_model: Qwen/Qwen2.5-VL-7B-Instruct\n\nchat_template: qwen2_vl  # same as qwen2-vl",
    "crumbs": [
      "How To Guides",
      "MultiModal / Vision Language Models (BETA)"
    ]
  },
  {
    "objectID": "docs/multimodal.html#dataset-format",
    "href": "docs/multimodal.html#dataset-format",
    "title": "MultiModal / Vision Language Models (BETA)",
    "section": "Dataset Format",
    "text": "Dataset Format\nFor multi-modal datasets, we adopt an extended chat_template format similar to OpenAI’s Message format.\n\nA message is a list of role and content.\nrole can be system, user, assistant, etc.\ncontent is a list of type and (text or image or path or url or base64).\n\n\n\n\n\n\n\nNote\n\n\n\nFor backwards compatibility:\n\nIf the dataset has a images or image column of list[Image], it will be appended to the first content list as {\"type\": \"image\", \"image\": ...}. However, if the content already has a {\"type\": \"image\"} but no image key, it will be set the image key.\nIf content is a string, it will be converted to a list with type as text.\n\n\n\n\n\n\n\n\n\nTip\n\n\n\nFor image loading, you can use the following keys within content alongside \"type\": \"image\":\n\n\"path\": \"/path/to/image.jpg\"\n\"url\": \"https://example.com/image.jpg\"\n\"base64\": \"...\"\n\"image\": PIL.Image\n\n\n\nHere is an example of a multi-modal dataset:\n[\n  {\n    \"messages\": [\n        {\n            \"role\": \"system\",\n            \"content\": [\n              {\"type\": \"text\", \"text\": \"You are a helpful assistant.\"}\n              ]\n        },\n        {\n            \"role\": \"user\",\n            \"content\": [\n                {\"type\": \"image\", \"url\": \"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg\"},\n                {\"type\": \"text\", \"text\": \"Describe this image in detail.\"}\n            ]\n        },\n        {\n            \"role\": \"assistant\",\n            \"content\": [\n              {\"type\": \"text\", \"text\": \"The image is a bee.\"}\n            ]\n        }\n    ]\n  }\n]",
    "crumbs": [
      "How To Guides",
      "MultiModal / Vision Language Models (BETA)"
    ]
  },
  {
    "objectID": "docs/reward_modelling.html",
    "href": "docs/reward_modelling.html",
    "title": "Reward Modelling",
    "section": "",
    "text": "Overview\nReward modelling is a technique used to train models to predict the reward or value of a given input. This is particularly useful in reinforcement learning scenarios where the model needs to evaluate the quality of its actions or predictions.\nWe support the reward modelling techniques supported by trl.\n\n\n(Outcome) Reward Models\nOutcome reward models are trained using data which contains preference annotations for an entire interaction between the user and model (e.g. rather than per-turn or per-step).\nbase_model: google/gemma-2-2b\nmodel_type: AutoModelForSequenceClassification\nnum_labels: 1\ntokenizer_type: AutoTokenizer\n\nreward_model: true\nchat_template: gemma\ndatasets:\n  - path: argilla/distilabel-intel-orca-dpo-pairs\n    type: bradley_terry.chat_template\n\nval_set_size: 0.1\neval_steps: 100\nBradley-Terry chat templates expect single-turn conversations in the following format:\n{\n    \"system\": \"...\", // optional\n    \"input\": \"...\",\n    \"chosen\": \"...\",\n    \"rejected\": \"...\"\n}\n\n\nProcess Reward Models (PRM)\n\n\n\n\n\n\nTip\n\n\n\nCheck out our PRM blog.\n\n\nProcess reward models are trained using data which contains preference annotations for each step in a series of interactions. Typically, PRMs are trained to provide reward signals over each step of a reasoning trace and are used for downstream reinforcement learning.\nbase_model: Qwen/Qwen2.5-3B\nmodel_type: AutoModelForTokenClassification\nnum_labels: 2\n\nprocess_reward_model: true\ndatasets:\n  - path: trl-lib/math_shepherd\n    type: stepwise_supervised\n    split: train\n\nval_set_size: 0.1\neval_steps: 100\nPlease see stepwise_supervised for more details on the dataset format.",
    "crumbs": [
      "How To Guides",
      "Reward Modelling"
    ]
  },
  {
    "objectID": "docs/faq.html",
    "href": "docs/faq.html",
    "title": "FAQ",
    "section": "",
    "text": "General\nQ: The trainer stopped and hasn’t progressed in several minutes.\n\nA: Usually an issue with the GPUs communicating with each other. See the NCCL doc\n\nQ: Exitcode -9\n\nA: This usually happens when you run out of system RAM.\n\nQ: Exitcode -7 while using deepspeed\n\nA: Try upgrading deepspeed w: pip install -U deepspeed\n\nQ: AttributeError: ‘DummyOptim’ object has no attribute ‘step’\nQ: ModuleNotFoundError: No module named ‘mpi4py’ using single GPU with deepspeed\n\nA: You may be using deepspeed with single gpu. Please remove the deepspeed: section in the yaml file or --deepspeed CLI flag.\n\nQ: The codes is stuck on saving preprocessed datasets.\n\nA: This is usually an issue with the GPU. This can be resolved through setting the os environment variable CUDA_VISIBLE_DEVICES=0. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.\n\nQ: Received mismatch error on merge adapters / loading adapters between torch.Size of checkpoint and model.\n\nA: This is likely due to vocab size mismatch. By default, Axolotl expands the model’s embeddings if the tokenizer has more tokens than the model. Please use the axolotl merge-lora command to merge the adapters instead of using your own scripts.\n\n\nOn the other hand, if the model has more tokens than the tokenizer, Axolotl does not shrink the model’s embeddings unless shrink_embeddings: true is set in the config.\n\nQ: How to call Axolotl via custom python scripts?\n\nA: Since Axolotl is just Python, please see src/axolotl/cli/main.py on how each command is called.\n\nQ: How to know the value to use for fsdp_transformer_layer_cls_to_wrap?\n\nA: This is the class name of the transformer layer to wrap with FSDP. For example, for LlamaForCausalLM, the value is LlamaDecoderLayer. To find this for a specific model, check the model’s PreTrainedModel definition and look for _no_split_modules variable in the modeling_<model_name>.py file within transformers library.\n\nQ: ValueError: Asking to pad but the tokenizer does not have a padding token. Please select a token to use as pad_token\n\nA: This is because the tokenizer does not have a padding token. Please add a padding token to the tokenizer via:\n\n\nspecial_tokens:\n  # str. If you're not sure, set to same as `eos_token`.\n  pad_token: \"...\"\n\n\n\nChat templates\nQ: jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____\n\nA: This means that the property mapping for the stated attribute does not exist when building chat_template prompt. For example, if no attribute 'content', please check you have added the correct mapping for content under message_property_mappings.\n\nQ: Empty template generated for turn ___\n\nA: The content is empty for that turn.\n\nQ: Could not find content start/end boundary for turn __\n\nA: The specific turn’s start/end could not be detected. Please ensure you have set the eos_token following your chat_template. Otherwise, this could be a chat_template which doesn’t use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not [[dummy_message]]. Please let us know about this.\n\nQ: Content end boundary is before start boundary for turn ___\n\nA: This is an edge case which should not occur. Please create an Issue if this happens.\n\nQ: Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.\n\nA: This is likely an empty turn.\n\nQ: The EOS token is incorrectly being masked or not being masked / EOS token __ not found in chat template.\n\nA: There can be two reasons:\n\n\n\nThis is because of the mismatch between tokenizer.eos_token and EOS token in template. Please make sure to set eos_token: under special_tokens: to the same EOS token as in template.\n\n\n\n\nThe EOS token is not in the template. Please check if your template is correct. As an example, phi_35 template does not use its dedicated EOS token <|endoftext|> at the end.\n\n\nQ: “chat_template choice is tokenizer_default but tokenizer’s chat_template is null. Please add a chat_template in tokenizer config”\n\nA: This is because the tokenizer does not have a chat template. Please add a chat template in the tokenizer config. See chat_template for more details.\n\nQ: The EOT token(s) are incorrectly being masked or not being masked / EOT token __ not found in chat template.\n\nA: There can be two reasons:\n\n\n\nThe EOT token is different from the EOS token and was not specified under eot_tokens:. Please set eot_tokens: to the same EOT token(s) as in template.\n\n\n\n\nThere is more than one EOT token per turn in the template. Please raise an issue with examples as we recognize this as an edge case.\n\n\nQ: EOT token encoding failed. Please check if the token is valid and can be encoded.\n\nA: There could be some issue with the tokenizer or unicode encoding. Please raise an issue with examples with the EOT token & tokenizer causing the issue.\n\nQ: EOT token __ is encoded as multiple tokens.\n\nA: This is because the EOT token is encoded as multiple tokens which can cause unexpected behavior. Please add it under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:.\n\nQ: Conflict between train_on_eos and train_on_eot. eos_token is in eot_tokens and train_on_eos != train_on_eot\n\nA: This is because the EOS token is in the eot_tokens: while mismatch between train_on_eos: and train_on_eot:. This will cause one to override the other. Please ensure that train_on_eos: and train_on_eot: are the same or remove the EOS token from eot_tokens:.\n\nQ: If eot_tokens: is not provided, what happens?\n\nA: If eot_tokens: is not provided, the default behavior is the same as before. EOS tokens used to delimit turns are masked/unmasked depending on whether the turn is trainable.\n\n\nInternally, eot_tokens: tokenizer.eos_token and train_on_eot: train_on_eos (which defaults to turn). This transition helps clarify the naming and behavior of EOT/EOS tokens.\n\nQ: Data processing error: CAS service error\n\nA: Try disabling XET with export HF_HUB_DISABLE_XET=1\n\nQ: torch._inductor.exc.LoweringException: NoValidChoicesError: No choices to select, please consider adding ATEN into max_autotune_gemm_backends config (defined in torch/_inductor/config.py) to allow at least one choice.\n\nA: Depending on the version of torch, you may need to include this in your YAML:\n\n\nflex_attn_compile_kwargs:\n  dynamic: false\n  mode: max-autotune-no-cudagraphs",
    "crumbs": [
      "Troubleshooting",
      "FAQ"
    ]
  },
  {
    "objectID": "docs/installation.html",
    "href": "docs/installation.html",
    "title": "Installation",
    "section": "",
    "text": "This guide covers all the ways you can install and set up Axolotl for your environment.",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/installation.html#sec-requirements",
    "href": "docs/installation.html#sec-requirements",
    "title": "Installation",
    "section": "1 Requirements",
    "text": "1 Requirements\n\nNVIDIA GPU (Ampere architecture or newer for bf16 and Flash Attention) or AMD GPU\nPython ≥3.10\nPyTorch ≥2.5.1",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/installation.html#sec-installation-methods",
    "href": "docs/installation.html#sec-installation-methods",
    "title": "Installation",
    "section": "2 Installation Methods",
    "text": "2 Installation Methods\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure to have Pytorch installed before installing Axolotl in your local environment.\nFollow the instructions at: https://pytorch.org/get-started/locally/\n\n\n\n\n\n\n\n\nImportant\n\n\n\nFor Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.\n\n\n\n2.1 PyPI Installation (Recommended)\npip3 install -U packaging setuptools wheel ninja\npip3 install --no-build-isolation axolotl[flash-attn,deepspeed]\nWe use --no-build-isolation in order to detect the installed PyTorch version (if\ninstalled) in order not to clobber it, and so that we set the correct version of\ndependencies that are specific to the PyTorch version or other installed\nco-dependencies.\n\n\n2.2 uv Installation\nuv is a fast, reliable Python package installer and resolver built in Rust. It offers significant performance improvements over pip and provides better dependency resolution, making it an excellent choice for complex environments.\nInstall uv if not already installed\ncurl -LsSf https://astral.sh/uv/install.sh | sh\nsource $HOME/.local/bin/env\nChoose your CUDA version to use with PyTorch; e.g. cu124, cu126, cu128,\nthen create the venv and activate\nexport UV_TORCH_BACKEND=cu126\nuv venv --no-project --relocatable\nsource .venv/bin/activate\nInstall PyTorch\n- PyTorch 2.6.0 recommended\nuv pip install packaging setuptools wheel\nuv pip install torch==2.6.0\nuv pip install awscli pydantic\nInstall axolotl from PyPi\nuv pip install --no-build-isolation axolotl[deepspeed,flash-attn]\n\n# optionally install with vLLM if you're using torch==2.6.0 and want to train w/ GRPO\nuv pip install --no-build-isolation axolotl[deepspeed,flash-attn,vllm]\n\n\n2.3 Edge/Development Build\nFor the latest features between releases:\ngit clone https://github.com/axolotl-ai-cloud/axolotl.git\ncd axolotl\npip3 install -U packaging setuptools wheel ninja\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\n2.4 Docker\ndocker run --gpus '\"all\"' --rm -it axolotlai/axolotl:main-latest\nFor development with Docker:\ndocker compose up -d\n\n\n\n\n\n\nAdvanced Docker Configuration\n\n\n\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it \\\n  --name axolotl --ipc=host \\\n  --ulimit memlock=-1 --ulimit stack=67108864 \\\n  --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl \\\n  -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \\\n  axolotlai/axolotl:main-latest\n\n\n\n\n\n\n\n\nImportant\n\n\n\nFor Blackwell GPUs, please use axolotlai/axolotl:main-py3.11-cu128-2.7.0 or the cloud variant axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0.\n\n\nPlease refer to the Docker documentation for more information on the different Docker images that are available.",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/installation.html#sec-cloud",
    "href": "docs/installation.html#sec-cloud",
    "title": "Installation",
    "section": "3 Cloud Environments",
    "text": "3 Cloud Environments\n\n3.1 Cloud GPU Providers\nFor providers supporting Docker:\n\nUse axolotlai/axolotl-cloud:main-latest\nAvailable on:\n\nLatitude.sh\nJarvisLabs.ai\nRunPod\nNovita\n\n\n\n\n3.2 Google Colab\nUse our example notebook.",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/installation.html#sec-platform-specific",
    "href": "docs/installation.html#sec-platform-specific",
    "title": "Installation",
    "section": "4 Platform-Specific Instructions",
    "text": "4 Platform-Specific Instructions\n\n4.1 macOS\npip3 install --no-build-isolation -e '.'\nSee Section 6 for Mac-specific issues.\n\n\n4.2 Windows\n\n\n\n\n\n\nImportant\n\n\n\nWe recommend using WSL2 (Windows Subsystem for Linux) or Docker.",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/installation.html#sec-env-managers",
    "href": "docs/installation.html#sec-env-managers",
    "title": "Installation",
    "section": "5 Environment Managers",
    "text": "5 Environment Managers\n\n5.1 Conda/Pip venv\n\nInstall Python ≥3.10\nInstall PyTorch: https://pytorch.org/get-started/locally/\nInstall Axolotl:\npip3 install -U packaging setuptools wheel ninja\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n(Optional) Login to Hugging Face:\nhuggingface-cli login",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/installation.html#sec-troubleshooting",
    "href": "docs/installation.html#sec-troubleshooting",
    "title": "Installation",
    "section": "6 Troubleshooting",
    "text": "6 Troubleshooting\nIf you encounter installation issues, see our FAQ and Debugging Guide.",
    "crumbs": [
      "Getting Started",
      "Installation"
    ]
  },
  {
    "objectID": "docs/unsloth.html",
    "href": "docs/unsloth.html",
    "title": "Unsloth",
    "section": "",
    "text": "Overview\nUnsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over\nstandard industry baselines.\n\n\n\n\n\n\nImportant\n\n\n\nDue to breaking changes in transformers v4.48.0, users will need to downgrade to <=v4.47.1 to use this patch.\nThis will later be deprecated in favor of LoRA Optimizations.\n\n\n\n\nInstallation\nThe following will install the correct unsloth and extras from source.\npython scripts/unsloth_install.py | sh\n\n\nUsage\nAxolotl exposes a few configuration options to try out unsloth and get most of the performance gains.\nOur unsloth integration is currently limited to the following model architectures:\n- llama\nThese options are specific to LoRA finetuning and cannot be used for multi-GPU finetuning\nunsloth_lora_mlp: true\nunsloth_lora_qkv: true\nunsloth_lora_o: true\nThese options are composable and can be used with multi-gpu finetuning\nunsloth_cross_entropy_loss: true\nunsloth_rms_norm: true\nunsloth_rope: true\n\n\nLimitations\n\nSingle GPU only; e.g. no multi-gpu support\nNo deepspeed or FSDP support (requires multi-gpu)\nLoRA + QLoRA support only. No full fine tunes or fp8 support.\nLimited model architecture support. Llama, Phi, Gemma, Mistral only\nNo MoE support.",
    "crumbs": [
      "Advanced Features",
      "Unsloth"
    ]
  },
  {
    "objectID": "docs/config.html",
    "href": "docs/config.html",
    "title": "Config Reference",
    "section": "",
    "text": "# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files\n# This can also be a relative path to a model on disk\nbase_model: ./llama-7b-hf\n# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)\nbase_model_ignore_patterns:\n# If the base_model repo on hf hub doesn't include configuration .json files,\n# You can set that here, or leave this empty to default to base_model\nbase_model_config: ./llama-7b-hf\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model:\n# Optional tokenizer configuration path in case you want to use a different tokenizer\n# than the one defined in the base model\ntokenizer_config:\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too\nmodel_type: AutoModelForCausalLM\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: AutoTokenizer\n# Trust remote code for untrusted source\ntrust_remote_code:\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast:\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy:\n# Resize the model embeddings when new tokens are added to multiples of 32\n# This is reported to improve training speed on some models\nresize_token_embeddings_to_32x:\n# Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.\nshrink_embeddings:\n# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs\nembeddings_skip_upcast:\n# Whether to load the model with randomly initialized weights. Useful for\n# pre-training a model from scratch or debugging purposes.\nrandom_init_weights:\n\n# (Internal use only)\n# Used to identify which the model is based on\nis_falcon_derived_model:\nis_llama_derived_model:\nis_qwen_derived_model:\n# Please note that if you set this to true, `padding_side` will be set to \"left\" by default\nis_mistral_derived_model:\n\n# optional overrides to the base model configuration\noverrides_of_model_config:\n  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653\n  rope_scaling:\n    type: # linear | dynamic\n    factor: # float\n\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs:\n  # use_cache: False\n\n# optional overrides to the bnb 4bit quantization configuration\n# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig\nbnb_config_kwargs:\n  # These are default values\n  llm_int8_has_fp16_weight: false\n  bnb_4bit_quant_type: nf4\n  bnb_4bit_use_double_quant: true\n\n# quantization aware training\nqat:\n  activation_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for activation quantization. Valid options are \"int4\" and \"int8\"\n  weight_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for weight quantization. Valid options are \"int4\" and \"int8\"\n  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization\n  fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after\n\n# post-training quantization\nquantization:\n  weight_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8\n  activation_dtype: # Optional[str] = \"int8\". Fake quantization layout to use for activation quantization. Valid options are \"int4\" and \"int8\"\n  group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization\n  quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer.\n\n\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: true\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: true\n# Use bitsandbytes 4 bit\nload_in_4bit:\n\n# Use CUDA bf16\nbf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere\n# Use CUDA fp16\nfp16: true\n# Use CUDA tf32\ntf32: true # require >=ampere\n# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting\n\n# No AMP (automatic mixed precision)\nbfloat16: true # require >=ampere\nfloat16: true\n\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset\ngpu_memory_limit: 20GiB\n# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge\nlora_on_cpu: true\n\n# List[str]. Add plugins to extend the pipeline.\n# See `src/axolotl/integrations` for the available plugins or doc below for more details.\n# https://docs.axolotl.ai/docs/custom_integrations.html\nplugins:\n  # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n# A list of one or more datasets to finetune the model with\n# See https://docs.axolotl.ai/docs/dataset_loading.html for guide on loading datasets\n# See https://docs.axolotl.ai/docs/dataset-formats/ for guide on dataset formats\ndatasets:\n  # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n  - path: vicgalle/alpaca-gpt4\n    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n    type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>\n    ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file\n    data_files: # Optional[str] path to source data files\n\n    shards: # Optional[int] split dataset into N pieces (use with shards_idx)\n    shards_idx: # Optional[int] = 0 the index of sharded dataset to use\n\n    preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)\n\n    name: # Optional[str] name of dataset configuration to load\n    split: train # Optional[str] name of dataset split to load from\n    revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.\n    trust_remote_code: # Optional[bool] Trust remote code for untrusted source\n\n  # Custom user instruction prompt\n  - path: repo\n    type:\n      # The below are defaults. only set what's needed if you use a different column name.\n      system_prompt: \"\"\n      system_format: \"{system}\"\n      field_system: system\n      field_instruction: instruction\n      field_input: input\n      field_output: output\n\n      # Customizable to be single line or multi-line\n      # Use {instruction}/{input} as key to be replaced\n      # 'format' can include {input}\n      format: |-\n        User: {instruction} {input}\n        Assistant:\n      # 'no_input_format' cannot include {input}\n      no_input_format: \"{instruction} \"\n\n      # For `completion` datsets only, uses the provided field instead of `text` column\n      field:\n\n  # Using chat template\n  - path: ...\n    # Set type to `chat_template` to use this strategy\n    type: chat_template\n    # Specify the name of the chat template to use\n    # The name of the chat template to use for training, following values are supported:\n    # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.\n    # - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py\n    # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.\n    # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n    chat_template: tokenizer_default\n\n    # Custom jinja chat template. Used only if `chat_template: jinja` or empty.\n    chat_template_jinja:\n\n    # Key containing the messages (default: \"messages\")\n    field_messages: messages\n\n    # Key containing the system message (default: \"system\")\n    # If the system message is not present in the dataset sample, it will be loaded from the field_system property.\n    field_system: system\n\n    # Mapping of properties from the input dataset to the chat template.\n    # (default: message_property_mappings={'role':'role', 'content':'content'})\n    # If a property exists in the template but not in this mapping, the system will attempt\n    # to load it directly from the message using the property name as the key.\n    # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',\n    # while 'value' is loaded and used as 'content' in the chat template.\n    message_property_mappings:\n      role: from\n      content: value\n      # ...\n\n    # Optional[Dict[str, List]]. Roles mapping in the messages.\n    # The format is {target_role: [source_roles]}. All source roles will be mapped to the target role.\n    # The default is:\n    roles:\n      user: [\"human\", \"user\"]\n      assistant: [\"gpt\", \"assistant\"]\n      system: [\"system\"]\n      tool: [\"tool\"]\n\n    # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template.\n    # This does not drop the default system message from chat_template if it exists. If you wish to,\n    # we recommend using a custom jinja template with the default system message removed or\n    # adding a system turn with empty content.\n    drop_system_message:\n\n    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags\n    # See example at `docs/dataset-formats/conversation.qmd`\n    split_thinking:\n\n    # IMPORTANT: The following fields determine which parts of the conversation to train on.\n    # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train\n    # See examples at `docs/dataset-formats/conversation.qmd`\n    # Note: If the below 5 fields are empty, defaults to training only on the last message.\n\n    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.\n    roles_to_train: [\"assistant\"]  # default\n    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:\n    # - all: train on all EOS tokens\n    # - turn (default): train on the EOS token at the end of each trainable turn\n    # - last: train on the last EOS token in the conversation\n    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.\n    train_on_eos: turn\n    # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are:\n    # - all: train on all EOT tokens\n    # - turn: train on the EOT token at the end of each trainable turn\n    # - last: train on the last EOT token in the conversation\n    # If not specified, defaults to the value of train_on_eos for backward compatibility.\n    train_on_eot:\n    # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.\n    message_field_training: training\n    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.\n    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).\n    message_field_training_detail: train_detail\n\n\n# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\n# Deduplicates datasets and test_datasets with identical entries.\ndataset_exact_deduplication: true\n\n# A list of one or more datasets to eval the model with.\n# You can use either test_datasets, or val_set_size, but not both.\ntest_datasets:\n  - path: /workspace/data/eval.jsonl\n    ds_type: json\n    # You need to specify a split. For \"json\" datasets the default split is called \"train\".\n    split: train\n    type: completion\n    data_files:\n      - /workspace/data/eval.jsonl\n\n# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'\nrl:\nrl_beta:  # Optional[float]. The beta parameter for the RL training.\n\n# dpo\ndpo_use_weighting:  # Optional[bool]. Whether to perform weighting.\nrpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper.\n\n# orpo\norpo_alpha: 0.1  # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping.\n\n# kto\nkto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss.\nkto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss.\n\n# simpo\ncpo_alpha: 1.0  # Weight of the BC regularizer\nsimpo_gamma: 0.5  # Target reward margin for the SimPO loss\n\n# grpo\ntrl:\n  use_vllm: # Optional[bool]. Whether to use VLLM for RL training.\n  vllm_server_host: # Optional[str]. Host of the vLLM server to connect to.\n  vllm_server_port: # Optional[int]. Port of the vLLM server to connect to.\n  vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.\n  vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding.\n\n  beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use\n  max_completion_length: # Optional[int]. Maximum length of the completion for RL training.\n\n  reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir.\n  reward_weights: # Optional[list[float]]. List of reward weights for the reward functions.\n\n  num_generations: # Optional[int]. Number of generations to sample.\n  log_completions: # Optional[bool]. Whether to log completions.\n  num_completions_to_print: # Optional[int]. Number of completions to print when log_completions is True.\n\n  sync_ref_model: # Optional[bool]. Whether to sync the reference model.\n  ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model.\n  ref_model_sync_steps: # Optional[int]. Sync steps for the reference model.\n  scale_rewards: # Optional[bool]. Whether to scale rewards by their standard deviation.\n\n  temperature: # Optional[float]. Sampling temperature for the GRPO policy.\n  top_p: # Optional[float]. Top-p sampling probability for the generation policy.\n  top_k: # Optional[int]. Top-k sampling for the generation policy.\n  min_p: # Optional[float]. Minimum probability for the generation policy.\n  repetition_penalty: # Optional[float]. Penalty for tokens that appear in prompt and generated text.\n\n  num_iterations: # Optional[int]. Number of iterations per batch (μ) for GRPO.\n  epsilon: # Optional[float]. Epsilon value for clipping in the GRPO algorithm.\n  epsilon_high: # Optional[float]. Upper-bound epsilon value for clipping in the GRPO algorithm.\n  use_liger_loss: # Optional[bool]. Whether to use Liger loss for GRPO.\n  loss_type: # Optional[str]. Loss formulation to use. Supported values: grpo, bnpo, dr_grpo.\n  mask_truncated_completions: # Optional[bool]. Whether to exclude truncated completions from loss calculation.\n\n\n# reward modelling: `True` or `False`\nreward_model:\n\n# process reward modelling: `True` or `False`\nprocess_reward_model:\n\n# The name of the chat template to use for training, following values are supported:\n# - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.\n# - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py\n# - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.\n# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n# The selected chat template will be saved to the tokenizer_config.json for easier inferencing\n# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.\nchat_template: tokenizer_default\n# custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.\nchat_template_jinja: null\n# Optional[List[str]]. Custom EOT (End-of-Turn) tokens to mask/unmask during training.\n# These tokens mark the boundaries between conversation turns.\n# For example: [\"/INST\", \"</s>\", \"[/SYSTEM_PROMPT]\"]\n# If not specified, defaults to just the model's eos_token.\n# This is useful for templates that use multiple delimiter tokens.\neot_tokens:\n  # - \"</s>\"\n  # - \"[/INST]\"\n  # - \"[/SYSTEM_PROMPT]\"\n# Changes the default system message\ndefault_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.\n# Axolotl attempts to save the dataset as an arrow after packing the data together so\n# subsequent training attempts load faster, relative path\ndataset_prepared_path: data/last_run_prepared\n# Push prepared dataset to hub\npush_dataset_to_hub: # Optional[str] repo_org/repo_name\n# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`\n# if not set.\ndataset_processes: # defaults to os.cpu_count() if not set\n# Keep dataset in memory while preprocessing\n# Only needed if cached dataset is taking too much storage\ndataset_keep_in_memory:\n# push checkpoints to hub\nhub_model_id: # private repo path to push finetuned model\n# how to push checkpoints to hub\n# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy\nhub_strategy:\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets\n# Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: # boolean\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.\nval_set_size: 0.04\n# Num shards for whole dataset\ndataset_shard_num:\n# Index of shard to use for whole dataset\ndataset_shard_idx:\n\n# The maximum length of an input to train with, this should typically be less than 2048\n# as most models have a token/context limit of 2048\nsequence_len: 2048\n# Pad inputs so each step uses constant sized buffers\n# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently\npad_to_sequence_len:\n# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'\nsample_packing:\n# Set to 'false' if getting errors during eval with sample_packing on.\neval_sample_packing:\n# You can set these packing optimizations AFTER starting a training at least once.\n# The trainer will provide recommended values for these values.\nsample_packing_eff_est:\ntotal_num_tokens:\n# Increasing the following values helps with packing, but usually only slightly (<%1.)\n# The number of samples packed at a time.\nsample_packing_group_size: 100000\n# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.\nsample_packing_bin_size: 200\nsample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially.\n\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation:\n\ncurriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening:\n\n# Passed through to transformers when loading the model when launched without accelerate\n# Use `sequential` when training w/ model parallelism to limit memory\ndevice_map:\n# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.\nmax_memory:\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model\nadapter: lora\n# If you already have a lora model trained that you want to load, put that here.\n# This means after training, if you want to test the model, you should set this to the value of `output_dir`.\n# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir:\n\n# LoRA hyperparameters\n# For more details about the following options, see:\n# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n  - q_proj\n  - v_proj\n#  - k_proj\n#  - o_proj\n#  - gate_proj\n#  - down_proj\n#  - up_proj\nlora_target_linear: # If true, will target all linear modules\n\n# List[int] | int. # The layer indices to transform, otherwise, apply to all layers\n# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform\npeft_layers_to_transform:\n\n# Optional[bool]. Whether to use DoRA.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora\npeft_use_dora:\n\n# Optional[bool]. Whether to use RSLoRA.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora\npeft_use_rslora:\n\n# Optional[list[tuple[int, int]]]. List of layer indices to replicate.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora\npeft_layer_replication:\n\n# bool | Literal[\"gaussian\", \"eva\", \"olora\", \"pissa\", \"pissa_niter_[number of iters]\", \"corda\", \"loftq\"]\n# How to initialize LoRA weights. Default to True which is MS original implementation.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization\npeft_init_lora_weights:\n\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.\n# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\n# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994\nlora_modules_to_save:\n#  - embed_tokens\n#  - lm_head\n\nlora_fan_in_fan_out: false\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for\n# speed and memory savings\n# See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true\n\n# LoRA+ hyperparameters\n# For more details about the following options, see:\n# https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`\nloraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_embedding: #  loraplus learning rate for lora embedding layers. Default value is 1e-6.\n\npeft:\n  # Configuration options for loftq initialization for LoRA\n  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization\n  loftq_config:\n    loftq_bits:  # typically 4 bits\n\n# ReLoRA configuration\n# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed\nrelora_steps: # Number of steps per ReLoRA restart\nrelora_warmup_steps: # Number of per-restart warmup steps\nrelora_anneal_steps: # Number of anneal steps for each relora cycle\nrelora_prune_ratio: # threshold for optimizer magnitude when pruning\nrelora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings\n\n# wandb configuration if you're using it\n# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.\nwandb_mode: # \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn off wandb\nwandb_project: # Your wandb project name\nwandb_entity: # A wandb Team name if using a Team\nwandb_watch:\nwandb_name: # Set the name of your wandb run\nwandb_run_id: # Set the ID of your wandb run\nwandb_log_model: # \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only at the end of training\n\n# mlflow configuration if you're using it\nmlflow_tracking_uri: # URI to mlflow\nmlflow_experiment_name: # Your experiment name\nmlflow_run_name: # Your run name\nhf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry\n\n# Comet configuration if you're using it\n# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.\n# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start\nuse_comet: # Enable or disable Comet integration.\ncomet_api_key: # API key for Comet. Recommended to set via `comet login`.\ncomet_workspace: # Workspace name in Comet. Defaults to the user's default workspace.\ncomet_project_name: # Project name in Comet. Defaults to Uncategorized.\ncomet_experiment_key: # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.\ncomet_mode: # Create a new experiment (\"create\") or log to an existing one (\"get\"). Default (\"get_or_create\") auto-selects based on configuration.\ncomet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.\ncomet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.\n\n# Tensorboard\nuse_tensorboard: # Optional[bool]\n\n# Where to save the full-finetuned model to\noutput_dir: ./completed-model\n\n# Whether to use torch.compile and which backend to use\n# setting to `auto` will enable torch compile when torch>=2.5.1\ntorch_compile:  # Optional[Union[Literal[\"auto\"], bool]]\ntorch_compile_backend:  # Optional[str]\ntorch_compile_mode:  # 'default' | 'reduce-overhead' | 'max-autotune'\n\n# Training hyperparameters\n\n# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.\ngradient_accumulation_steps: 1\n# The number of samples to include in each batch. This is the number of samples sent to each GPU.\n# Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: 2\neval_batch_size:\nnum_epochs: 4\nwarmup_steps: 100  # cannot use with warmup_ratio\nwarmup_ratio: 0.05  # cannot use with warmup_steps\nlearning_rate: 0.00003\nlr_quadratic_warmup:\nlogging_steps:\neval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\neval_strategy: # Set to `\"no\"` to skip evaluation, `\"epoch\"` at end of each epoch, leave empty to infer from `eval_steps`.\nsave_strategy: # Set to `\"no\"` to skip checkpoint saves, `\"epoch\"` at end of each epoch, `\"best\"` when better result is achieved, leave empty to infer from `save_steps`.\nsave_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps\nsaves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsave_total_limit: # Checkpoints saved at a time\nsave_only_model: # Save only the model weights, skipping the optimizer. Using this means you can't resume from checkpoints.\n# Maximum number of iterations to train for. It precedes num_epochs which means that\n# if both are set, num_epochs will not be guaranteed.\n# e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps:\n\n# bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.\ninclude_tokens_per_second: # Optional[bool]\n\n# whether to find batch size that fits in memory. Passed to underlying transformers Trainer\nauto_find_batch_size: # Optional[bool]\n\neval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0\neval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128\ndo_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`.\neval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is [\"sacrebleu\", \"comet\", \"ter\", \"chrf\", \"perplexity\"]\n\nprofiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.\n                # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information\n                # snapshots can be visualized @ https://pytorch.org/memory_viz\n\nloss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)\nloss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)\n\n# Save model as safetensors (require safetensors package). Default True\nsave_safetensors:\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: false\n# Group similarly sized data to minimize padding.\n# May be slower to start, as it must download and sort the entire dataset.\n# Note that training loss may have an oscillating pattern with this enabled.\ngroup_by_length: false\n\n# Whether to use gradient checkpointing. Available options are: true, false, \"offload\", \"offload_disk\".\n# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: false\n# additional kwargs to pass to the trainer for gradient checkpointing\n# gradient_checkpointing_kwargs:\n#   use_reentrant: true\n\n# Stop training after this many evaluation losses have increased in a row\n# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback\nearly_stopping_patience: 3\n\n# Specify a scheduler and kwargs to use with the optimizer\n# Valid values are driven by the Transformers SchedulerType class, see:\n# https://github.com/huggingface/transformers/blob/5f4ecf2d9f867a1255131d2461d75793c0cf1db2/src/transformers/trainer_utils.py#L420\n# Valid values include\n# - 'linear'\n# - 'cosine' (default)\n# - 'cosine_with_restarts'\n# - 'polynomial'\n# - 'constant'\n# - 'constant_with_warmup'\n# - 'inverse_sqrt'\n# - 'reduce_lr_on_plateau'\n# - 'cosine_with_min_lr'\n# - 'warmup_stable_decay'\n\n# Additional schedulers include:\n# - 'one_cycle'\n# - 'rex'\nlr_scheduler:\nlr_scheduler_kwargs:\ncosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr\ncosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)\n\n# For one_cycle optim\nlr_div_factor: # Learning rate div factor\n\n# Specify optimizer\n# Valid values are driven by the Transformers OptimizerNames class, see:\n# https://github.com/huggingface/transformers/blob/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189\n#\n# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of\n# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used\n# in the examples/ for your model and fine-tuning use case.\n#\n# Valid values for 'optimizer' include:\n# - adamw_torch\n# - adamw_torch_fused (default)\n# - adamw_torch_xla\n# - adamw_torch_npu_fused\n# - adamw_apex_fused\n# - adopt_adamw  (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)\n# - adafactor\n# - adamw_anyprecision\n# - adamw_torch_4bit\n# - ademamix\n# - sgd\n# - adagrad\n# - adamw_bnb_8bit\n# - adamw_8bit   # alias for adamw_bnb_8bit\n# - ademamix_8bit\n# - lion_8bit\n# - lion_32bit\n# - paged_adamw_32bit\n# - paged_adamw_8bit\n# - paged_ademamix_32bit\n# - paged_ademamix_8bit\n# - paged_lion_32bit\n# - paged_lion_8bit\n# - rmsprop\n# - rmsprop_bnb\n# - rmsprop_bnb_8bit\n# - rmsprop_bnb_32bit\n# - galore_adamw\n# - galore_adamw_8bit\n# - galore_adafactor\n# - galore_adamw_layerwise\n# - galore_adamw_8bit_layerwise\n# - galore_adafactor_layerwise\n# - lomo\n# - adalomo\n# - grokadamw\n# - schedule_free_adamw\n# - schedule_free_sgd\n# - apollo_adamw\n# - apollo_adamw_layerwise\n#\n# Additional custom optimizers include:\n# - optimi_adamw\n# - ao_adamw_8bit\n# - ao_adamw_fp8\n# - came_pytorch\noptimizer:\n# Dictionary of arguments to pass to the optimizer\noptim_args:\n# For Galore Optimizers the following optim_args are available\n# rank:  # type: int\n# update_proj_gap  # type: int\n# scale  # type: float\n# proj_type:  # type: str, default = std\n\n# The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm\noptim_target_modules:\n# - self_attn  # for llama\n# - mlp\n\n# Specify weight decay\nweight_decay:\n# adamw hyperparams\nadam_beta1:\nadam_beta2:\nadam_beta3:  # only used for CAME Optimizer\nadam_epsilon:\nadam_epsilon2:  # only used for CAME Optimizer\n# Gradient clipping max norm\nmax_grad_norm:\n\n# Augmentation techniques\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings\n# currently only supported on Llama and Mistral\nneftune_noise_alpha:\n\n# Optional[bool]. Whether to bettertransformers\nflash_optimum:\n\n# Note: Only one of the following attention patches can be used at a time.\n# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.\n\n# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:\nxformers_attention:\n# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:\nflash_attention:\nflash_attn_cross_entropy:  # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_rms_norm:  # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation\nflash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation\n# Optional[bool]. Whether to use scaled-dot-product attention\n# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention:\n# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention:\n\n# Optional[bool]. Whether to use low_cpu_mem_usage\nlow_cpu_mem_usage:\n# Optional[str]. Resume from a specific checkpoint dir\nresume_from_checkpoint:\n# Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off.\n# Be careful with this being turned on between different models.\nauto_resume_from_checkpoints: false\n\n## Multimodal section\n# int | tuple[int, int] | None . Size to resize images to, width x height.\n# Will read from model/processor config if not set.\nimage_size:\n# str. Algorithm to use for image resizing. \"bilinear\", \"bicubic\", \"lanczos\". Default is \"bilinear\".\nimage_resize_algorithm: 'bilinear'\n## End of multimodal section\n\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank:\n\n# Add or change special tokens.\n# If you add tokens here, you don't need to add them to the `tokens` list.\nspecial_tokens:\n  # bos_token: \"<s>\"\n  # eos_token: \"</s>\"\n  # unk_token: \"<unk>\"\n  # pad_token: \"[PAD]\"\n\n# Optional[list[str]]. Add extra tokens to the tokenizer.\ntokens:\n  # - \"<|startoftext|>\"\n  # - \"<|endoftext|>\"\n\n# Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.\n# Only works for tokens that are not part of the base vocab (aka are added_tokens).\n# Can be checked if they exist in tokenizer.json added_tokens.\nadded_tokens_overrides:  # Dict[int, str]\n#  128041: \"<|im_start|>\"\n#  128042: \"<|im_end|>\"\n\n# FSDP\nfsdp:\nfsdp_config:\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed:\n\n# Advanced DDP Arguments\nddp_timeout:\nddp_bucket_cap_mb:\nddp_broadcast_buffers:\n\n# Sequence parallelism\n# Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.\n# Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.\n# E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized\n# subsequences, or set to 4 to split into four equal-sized subsequences.\n# See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.\nsequence_parallel_degree:\n# Optional; strides across the key dimension. Larger values use more memory but should make training faster.\n# Must evenly divide the number of KV heads in your model.\nheads_k_stride: 1\n# One of \"varlen_llama3\", \"batch_ring\", \"batch_zigzag\", \"batch_stripe\". Defaults to \"varlen_llama3\"\n# in the sample packing case, and \"batch_ring\" in the non-sample packing case.\nring_attn_func:\n\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path:\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset:\n\n# Debug mode\ndebug:\n\n# Seed\nseed:\n\n# Allow overwrite yml config using from cli\nstrict:",
    "crumbs": [
      "Getting Started",
      "Config Reference"
    ]
  },
  {
    "objectID": "docs/inference.html",
    "href": "docs/inference.html",
    "title": "Inference and Merging",
    "section": "",
    "text": "This guide covers how to use your trained models for inference, including model loading, interactive testing, merging adapters, and common troubleshooting steps.",
    "crumbs": [
      "Getting Started",
      "Inference and Merging"
    ]
  },
  {
    "objectID": "docs/inference.html#sec-quickstart",
    "href": "docs/inference.html#sec-quickstart",
    "title": "Inference and Merging",
    "section": "1 Quick Start",
    "text": "1 Quick Start\n\n\n\n\n\n\nTip\n\n\n\nUse the same config used for training on inference/merging.\n\n\n\n1.1 Basic Inference\n\nLoRA ModelsFull Fine-tuned Models\n\n\naxolotl inference your_config.yml --lora-model-dir=\"./lora-output-dir\"\n\n\naxolotl inference your_config.yml --base-model=\"./completed-model\"",
    "crumbs": [
      "Getting Started",
      "Inference and Merging"
    ]
  },
  {
    "objectID": "docs/inference.html#sec-advanced",
    "href": "docs/inference.html#sec-advanced",
    "title": "Inference and Merging",
    "section": "2 Advanced Usage",
    "text": "2 Advanced Usage\n\n2.1 Gradio Interface\nLaunch an interactive web interface:\naxolotl inference your_config.yml --gradio\n\n\n2.2 File-based Prompts\nProcess prompts from a text file:\ncat /tmp/prompt.txt | axolotl inference your_config.yml \\\n  --base-model=\"./completed-model\" --prompter=None\n\n\n2.3 Memory Optimization\nFor large models or limited memory:\naxolotl inference your_config.yml --load-in-8bit=True",
    "crumbs": [
      "Getting Started",
      "Inference and Merging"
    ]
  },
  {
    "objectID": "docs/inference.html#sec-merging",
    "href": "docs/inference.html#sec-merging",
    "title": "Inference and Merging",
    "section": "3 Merging LoRA Weights",
    "text": "3 Merging LoRA Weights\nMerge LoRA adapters with the base model:\naxolotl merge-lora your_config.yml --lora-model-dir=\"./completed-model\"\n\n3.1 Memory Management for Merging\n\nConfiguration OptionsForce CPU Merging\n\n\ngpu_memory_limit: 20GiB  # Adjust based on your GPU\nlora_on_cpu: true        # Process on CPU if needed\n\n\nCUDA_VISIBLE_DEVICES=\"\" axolotl merge-lora ...",
    "crumbs": [
      "Getting Started",
      "Inference and Merging"
    ]
  },
  {
    "objectID": "docs/inference.html#sec-tokenization",
    "href": "docs/inference.html#sec-tokenization",
    "title": "Inference and Merging",
    "section": "4 Tokenization",
    "text": "4 Tokenization\n\n4.1 Common Issues\n\n\n\n\n\n\nWarning\n\n\n\nTokenization mismatches between training and inference are a common source of problems.\n\n\nTo debug:\n\nCheck training tokenization:\n\naxolotl preprocess your_config.yml --debug\n\nVerify inference tokenization by decoding tokens before model input\nCompare token IDs between training and inference\n\n\n\n4.2 Special Tokens\nConfigure special tokens in your YAML:\nspecial_tokens:\n  bos_token: \"<s>\"\n  eos_token: \"</s>\"\n  unk_token: \"<unk>\"\ntokens:\n  - \"<|im_start|>\"\n  - \"<|im_end|>\"",
    "crumbs": [
      "Getting Started",
      "Inference and Merging"
    ]
  },
  {
    "objectID": "docs/inference.html#sec-troubleshooting",
    "href": "docs/inference.html#sec-troubleshooting",
    "title": "Inference and Merging",
    "section": "5 Troubleshooting",
    "text": "5 Troubleshooting\n\n5.1 Common Problems\n\nMemory IssuesToken IssuesPerformance Issues\n\n\n\nUse 8-bit loading\nReduce batch sizes\nTry CPU offloading\n\n\n\n\nVerify special tokens\nCheck tokenizer settings\nCompare training and inference preprocessing\n\n\n\n\nVerify model loading\nCheck prompt formatting\nEnsure temperature/sampling settings\n\n\n\n\nFor more details, see our debugging guide.",
    "crumbs": [
      "Getting Started",
      "Inference and Merging"
    ]
  },
  {
    "objectID": "src/axolotl/integrations/LICENSE.html",
    "href": "src/axolotl/integrations/LICENSE.html",
    "title": "Axolotl",
    "section": "",
    "text": "AXOLOTL COMMUNITY LICENSE AGREEMENT\nThis Axolotl Community License Agreement (“Agreement”) is entered into by and between Axolotl AI Corp. (“Axolotl”) and\nany individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms\nand conditions set forth in this Agreement.\n\nDefinitions\n1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement.\n1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl,\nwhich may be licensed separately by their respective authors and/or licensors.\n1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at\nhttps://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which\npermits Plugin Integrations to integrate with the Axolotl service.\nGrant of License\n2.1 Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge,\npublish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions:\n- Licensee must comply with all the terms and conditions of this Agreement.\n- Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial\nportions of the Software.\n2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.\nRestrictions\n3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for\nfree or for sale any services, platform, or equivalent to third parties for the purposes of allowing such\nthird parties to fine-tune artificial intelligence models.\n3.2 Licensee shall not:\n- Use the Software for any illegal or unauthorized purpose.\n- Reverse engineer, decompile, or disassemble the Software.\n- Remove or modify any copyright, trademark, or other proprietary notices contained in the Software.\n- Use the Software in a way that could damage, disable, overburden, or impair the functionality of the\nSoftware or interfere with any third-party use of the Software.\n3.3 Axolotl reserves the right to restrict certain Plugin Integrations for use with the Software. To the extent Licensee integrates a permitted, applicable Plugin Integration with the Software, Licensee shall comply with any additional terms and conditions imposed by the licensors of such Plugin Integration for use of such Plugin Integrations. Licensee shall contact Axolotl if it has questions about whether its use of the Software falls beyond the scope of this Agreement.\nIntellectual Property Rights\n4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee\nacknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to\nLicensee.\nDisclaimer of Warranty\n5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED\nTO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL\nTHE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF\nCONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER\nDEALINGS IN THE SOFTWARE.\nTermination\n6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and\nconditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any\ncopies in its possession.\nGoverning Law\n7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California,\nwithout regards to conflicts of laws provisions thereof.\nEntire Agreement\n8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter\nhereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning\nthe Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and\nLicensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms\non a go-forward basis. Axolotl will use commercially reasonable efforts to provide Licensee notice of any\nmaterial updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be\nbound by the terms and conditions of this Agreement.\n\nThis Agreement was last updated on August 23, 2024."
  },
  {
    "objectID": "index.html",
    "href": "index.html",
    "title": "Axolotl",
    "section": "",
    "text": "Axolotl is a tool designed to streamline post-training for various AI models.\nPost-training refers to any modifications or additional training performed on\npre-trained models - including full model fine-tuning, parameter-efficient tuning (like\nLoRA and QLoRA), supervised fine-tuning (SFT), instruction tuning, and alignment\ntechniques. With support for multiple model architectures and training configurations,\nAxolotl makes it easy to get started with these techniques.\nAxolotl is designed to work with YAML config files that contain everything you need to\npreprocess a dataset, train or fine-tune a model, run model inference or evaluation,\nand much more.\nFeatures:",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#quick-start",
    "href": "index.html#quick-start",
    "title": "Axolotl",
    "section": "🚀 Quick Start",
    "text": "🚀 Quick Start\nRequirements:\n\nNVIDIA GPU (Ampere or newer for bf16 and Flash Attention) or AMD GPU\nPython 3.11\nPyTorch ≥2.5.1\n\n\nInstallation\npip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja\npip3 install --no-build-isolation axolotl[flash-attn,deepspeed]\n\n# Download example axolotl configs, deepspeed configs\naxolotl fetch examples\naxolotl fetch deepspeed_configs  # OPTIONAL\nOther installation approaches are described here.\n\n\nYour First Fine-tune\n# Fetch axolotl examples\naxolotl fetch examples\n\n# Or, specify a custom path\naxolotl fetch examples --dest path/to/folder\n\n# Train a model using LoRA\naxolotl train examples/llama-3/lora-1b.yml\nThat’s it! Check out our Getting Started Guide for a more detailed walkthrough.",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#key-features",
    "href": "index.html#key-features",
    "title": "Axolotl",
    "section": "✨ Key Features",
    "text": "✨ Key Features\n\nMultiple Model Support: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more\nTraining Methods: Full fine-tuning, LoRA, QLoRA, and more\nEasy Configuration: Simple YAML files to control your training setup\nPerformance Optimizations: Flash Attention, xformers, multi-GPU training\nFlexible Dataset Handling: Use various formats and custom datasets\nCloud Ready: Run on cloud platforms or local hardware",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#documentation",
    "href": "index.html#documentation",
    "title": "Axolotl",
    "section": "📚 Documentation",
    "text": "📚 Documentation\n\nInstallation Options - Detailed setup instructions for different environments\nConfiguration Guide - Full configuration options and examples\nDataset Guide - Supported formats and how to use them\nMulti-GPU Training\nMulti-Node Training\nMultipacking\nAPI Reference - Auto-generated code documentation\nFAQ - Frequently asked questions",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#getting-help",
    "href": "index.html#getting-help",
    "title": "Axolotl",
    "section": "🤝 Getting Help",
    "text": "🤝 Getting Help\n\nJoin our Discord community for support\nCheck out our Examples directory\nRead our Debugging Guide\nNeed dedicated support? Please contact ✉️wing@axolotl.ai for options",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#contributing",
    "href": "index.html#contributing",
    "title": "Axolotl",
    "section": "🌟 Contributing",
    "text": "🌟 Contributing\nContributions are welcome! Please see our Contributing Guide for details.",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#supported-models",
    "href": "index.html#supported-models",
    "title": "Axolotl",
    "section": "Supported Models",
    "text": "Supported Models\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nfp16/fp32\nlora\nqlora\ngptq\ngptq w/flash attn\nflash attn\nxformers attn\n\n\n\n\nllama\n✅\n✅\n✅\n✅\n✅\n✅\n✅\n\n\nMistral\n✅\n✅\n✅\n✅\n✅\n✅\n✅\n\n\nMixtral-MoE\n✅\n✅\n✅\n❓\n❓\n❓\n❓\n\n\nMixtral8X22\n✅\n✅\n✅\n❓\n❓\n❓\n❓\n\n\nPythia\n✅\n✅\n✅\n❌\n❌\n❌\n❓\n\n\ncerebras\n✅\n✅\n✅\n❌\n❌\n❌\n❓\n\n\nbtlm\n✅\n✅\n✅\n❌\n❌\n❌\n❓\n\n\nmpt\n✅\n❌\n❓\n❌\n❌\n❌\n❓\n\n\nfalcon\n✅\n✅\n✅\n❌\n❌\n❌\n❓\n\n\ngpt-j\n✅\n✅\n✅\n❌\n❌\n❓\n❓\n\n\nXGen\n✅\n❓\n✅\n❓\n❓\n❓\n✅\n\n\nphi\n✅\n✅\n✅\n❓\n❓\n❓\n❓\n\n\nRWKV\n✅\n❓\n❓\n❓\n❓\n❓\n❓\n\n\nQwen\n✅\n✅\n✅\n❓\n❓\n❓\n❓\n\n\nGemma\n✅\n✅\n✅\n❓\n❓\n✅\n❓\n\n\nJamba\n✅\n✅\n✅\n❓\n❓\n✅\n❓\n\n\n\n✅: supported\n❌: not supported\n❓: untested",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#sponsors",
    "href": "index.html#sponsors",
    "title": "Axolotl",
    "section": "❤️ Sponsors",
    "text": "❤️ Sponsors\nThank you to our sponsors who help make Axolotl possible:\n\nModal - Modal lets you run\njobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale,\nfine-tune large language models, run protein folding simulations, and much more.\n\nInterested in sponsoring? Contact us at wing@axolotl.ai",
    "crumbs": [
      "Home"
    ]
  },
  {
    "objectID": "index.html#license",
    "href": "index.html#license",
    "title": "Axolotl",
    "section": "📜 License",
    "text": "📜 License\nThis project is licensed under the Apache 2.0 License - see the LICENSE file for details.",
    "crumbs": [
      "Home"
    ]
  }
]