tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity
Files
77764de0dd5002c922bd511c08490bd70ce557ca
axolotl/search.json
Quarto GHA Workflow Runner 77764de0dd Built site for gh-pages
2025-02-13 21:02:41 +00:00

1319 lines
190 KiB
JSON
Raw Blame History

This file contains invisible Unicode characters
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
[
{
"objectID": "index.html",
"href": "index.html",
"title": "Axolotl",
"section": "",
"text": "🚀 Quick Start\n \n Installation\n Your First Fine-tune\n \n ✨ Key Features\n 📚 Documentation\n 🤝 Getting Help\n 🌟 Contributing\n Supported Models\n ❤️ Sponsors\n 📜 License\nAxolotl is a tool designed to streamline post-training for various AI models. Post-training refers to any modifications or additional training performed on pre-trained models - including full model fine-tuning, parameter-efficient tuning (like LoRA and QLoRA), supervised fine-tuning (SFT), instruction tuning, and alignment techniques. With support for multiple model architectures and training configurations, Axolotl makes it easy to get started with these techniques.\nAxolotl is designed to work with YAML config files that contain everything you need to preprocess a dataset, train or fine-tune a model, run model inference or evaluation, and much more.\nFeatures:",
"crumbs": [
"Home"
]
},
{
"objectID": "index.html#quick-start",
"href": "index.html#quick-start",
"title": "Axolotl",
"section": "🚀 Quick Start",
"text": "🚀 Quick Start\nRequirements: - NVIDIA GPU (Ampere or newer for bf16 and Flash Attention) or AMD GPU - Python 3.11 - PyTorch ≥2.4.1\n\nInstallation\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\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 ❌: not supported ❓: 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 jobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale, fine-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"
]
},
{
"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": [
"How-To Guides",
"NCCL"
]
},
{
"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.\n\n\n\n\n\n\nImportant\n\n\n\nEach machine needs a copy of Axolotl, we suggest using the same commit to ensure compatibility.\nYou will also need to have the same configuration file for your model on each machine.\nMake sure the main machine is reachable by other machines.\n\n\n\nAccelerate\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.\n\n\nRaytrain\nPlease see ray train doc here.\n\n\nTorchrun\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": [
"How-To Guides",
"Multi Node"
]
},
{
"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 construct a 4d attention mask. We only need to concatenate the sequences into a single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4, each character represents 256 tokens X 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: [[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
"crumbs": [
"How-To Guides",
"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 construct a 4d attention mask. We only need to concatenate the sequences into a single batch and let flash attention know where each new sequence begins.\n4k context, bsz =4, each character represents 256 tokens X 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: [[ 0, 11, 17, 24, 28, 36, 41 44, 48, 51, 55, 60, 64]]",
"crumbs": [
"How-To Guides",
"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 efficiency as we are not able to join multiple batches into a single batch due to context length limits without flash attention. We can use either Pytorch’s Scaled Dot Product Attention implementation or native Pytorch attention implementation along with 4d attention masks to pack sequences together and avoid cross attention.",
"crumbs": [
"How-To Guides",
"Multipack (Sample Packing)"
]
},
{
"objectID": "docs/config.html",
"href": "docs/config.html",
"title": "Config options",
"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\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\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`. require >=ampere\n# Use CUDA fp16\nfp16: true\n# Use CUDA tf32\ntf32: true # require >=ampere\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# A list of one or more datasets to finetune the model with\ndatasets:\n # HuggingFace dataset repo | s3://,gs:// path | \"json\" for local dataset, make sure to fill data_files\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 shards: # Optional[int] number of shards to split data into\n name: # Optional[str] name of dataset configuration to load\n train_on_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 # Key for role in each message (default: \"role\")\n message_field_role: role\n # Key for content in each message (default: \"content\")\n message_field_content: content\n\n # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:\n roles:\n user: [\"human\", \"user\"]\n assistant: [\"gpt\", \"assistant\"]\n system: [\"system\"]\n tool: [\"tool\"]\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 4 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 train_on_eos: last\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\nDeduplicates 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'\nrl:\n# whether to perform weighting if doing DPO training. Boolean.\ndpo_use_weighting:\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# 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: # repo path\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\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation:\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\npeft_layers_to_transform: # The layer indices to transform, otherwise, apply to all layers\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# 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# 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]\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\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\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\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)\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 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\nlr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine\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/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134\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_hf\n# - adamw_torch\n# - adamw_torch_fused\n# - adamw_torch_xla\n# - adamw_apex_fused\n# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)\n# - adafactor\n# - adamw_anyprecision\n# - sgd\n# - adagrad\n# - adamw_bnb_8bit\n# - lion_8bit\n# - lion_32bit\n# - paged_adamw_32bit\n# - paged_adamw_8bit\n# - paged_lion_32bit\n# - paged_lion_8bit\n# - galore_adamw\n# - galore_adamw_8bit\n# - galore_adafactor\n# - galore_adamw_layerwise\n# - galore_adamw_8bit_layerwise\n# - galore_adafactor_layerwise\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_epsilon:\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# Whether to bettertransformers\nflash_optimum:\n# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:\nxformers_attention:\n# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:\nflash_attention:\nflash_attn_cross_entropy: # Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_rms_norm: # Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_fuse_qkv: # Whether to fuse QKV into a single operation\nflash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation\n# 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# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention:\n# Resume from a specific checkpoint dir\nresume_from_checkpoint:\n# 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# 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# Add extra tokens.\ntokens:\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# 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": [
"Reference",
"Config options"
]
},
{
"objectID": "docs/getting-started.html",
"href": "docs/getting-started.html",
"title": "Getting Started with Axolotl",
"section": "",
"text": "This guide will walk you through your first model fine-tuning project with Axolotl.",
"crumbs": [
"How-To Guides",
"Getting Started with Axolotl"
]
},
{
"objectID": "docs/getting-started.html#sec-quick-example",
"href": "docs/getting-started.html#sec-quick-example",
"title": "Getting Started with Axolotl",
"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. Assuming 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": [
"How-To Guides",
"Getting Started with Axolotl"
]
},
{
"objectID": "docs/getting-started.html#sec-understanding",
"href": "docs/getting-started.html#sec-understanding",
"title": "Getting Started with Axolotl",
"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# hub_model_id: username/custom_model_name\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\nadapter: lora\nlora_model_dir:\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 LoRA adapter layers\nLoads and processes the dataset\nRuns the training loop\nSaves the trained model and / or LoRA weights",
"crumbs": [
"How-To Guides",
"Getting Started with Axolotl"
]
},
{
"objectID": "docs/getting-started.html#sec-custom",
"href": "docs/getting-started.html#sec-custom",
"title": "Getting Started with Axolotl",
"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\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 the 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 format them.\n\nPrepare your JSONL data in the specified format (in this case, the expected `alpaca format):\n\n{\"instruction\": \"Classify this text\", \"input\": \"I love this!\", \"output\": \"positive\"}\n{\"instruction\": \"Classify this text\", \"input\": \"Not good at all\", \"output\": \"negative\"}\nPlease consult the supported Dataset Formats for more details.\n\nRun the training:\n\naxolotl train my_training.yml",
"crumbs": [
"How-To Guides",
"Getting Started with Axolotl"
]
},
{
"objectID": "docs/getting-started.html#sec-common-tasks",
"href": "docs/getting-started.html#sec-common-tasks",
"title": "Getting Started with Axolotl",
"section": "4 Common Tasks",
"text": "4 Common Tasks\n\n4.1 Testing Your Model\nAfter training, test your model:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\"\n\n\n4.2 Preprocessing Data\nFor large datasets, preprocess first:\naxolotl preprocess my_training.yml\n\n\n4.3 Using a UI\nLaunch a Gradio interface:\naxolotl inference my_training.yml --lora-model-dir=\"./outputs/lora-out\" --gradio",
"crumbs": [
"How-To Guides",
"Getting Started with Axolotl"
]
},
{
"objectID": "docs/getting-started.html#sec-next-steps",
"href": "docs/getting-started.html#sec-next-steps",
"title": "Getting Started with Axolotl",
"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 Formats - Working with different data formats\nMulti-GPU Training\nMulti-Node Training",
"crumbs": [
"How-To Guides",
"Getting Started with Axolotl"
]
},
{
"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: Micro batch size: 3 Gradient accumulation steps: 2 Number of GPUs: 3 Total 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: Micro batch size: 2 Gradient accumulation steps: 1 Number of GPUs: 3 Total 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)"
},
{
"objectID": "docs/multi-gpu.html",
"href": "docs/multi-gpu.html",
"title": "Multi-GPU Training Guide",
"section": "",
"text": "This guide covers advanced training configurations for multi-GPU setups using Axolotl.",
"crumbs": [
"How-To Guides",
"Multi-GPU Training Guide"
]
},
{
"objectID": "docs/multi-gpu.html#sec-overview",
"href": "docs/multi-gpu.html#sec-overview",
"title": "Multi-GPU Training Guide",
"section": "1 Overview",
"text": "1 Overview\nAxolotl supports several methods for multi-GPU training:\n\nDeepSpeed (recommended)\nFSDP (Fully Sharded Data Parallel)\nFSDP + QLoRA",
"crumbs": [
"How-To Guides",
"Multi-GPU Training Guide"
]
},
{
"objectID": "docs/multi-gpu.html#sec-deepspeed",
"href": "docs/multi-gpu.html#sec-deepspeed",
"title": "Multi-GPU Training Guide",
"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\naccelerate launch -m axolotl.cli.train examples/llama-2/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 2 (zero2.json)\nZeRO Stage 3 (zero3.json)\n\nChoose based on your memory requirements and performance needs.",
"crumbs": [
"How-To Guides",
"Multi-GPU Training Guide"
]
},
{
"objectID": "docs/multi-gpu.html#sec-fsdp",
"href": "docs/multi-gpu.html#sec-fsdp",
"title": "Multi-GPU Training Guide",
"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\n\n\n3.2 FSDP + QLoRA\nFor combining FSDP with QLoRA, see our dedicated guide.",
"crumbs": [
"How-To Guides",
"Multi-GPU Training Guide"
]
},
{
"objectID": "docs/multi-gpu.html#sec-performance",
"href": "docs/multi-gpu.html#sec-performance",
"title": "Multi-GPU Training Guide",
"section": "4 Performance Optimization",
"text": "4 Performance Optimization\n\n4.1 Liger Kernel Integration\n\n\n\n\n\n\nNote\n\n\n\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\n\n\nConfiguration:\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",
"crumbs": [
"How-To Guides",
"Multi-GPU Training Guide"
]
},
{
"objectID": "docs/multi-gpu.html#sec-troubleshooting",
"href": "docs/multi-gpu.html#sec-troubleshooting",
"title": "Multi-GPU Training Guide",
"section": "5 Troubleshooting",
"text": "5 Troubleshooting\n\n5.1 NCCL Issues\nFor NCCL-related problems, see our NCCL troubleshooting guide.\n\n\n5.2 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": [
"How-To Guides",
"Multi-GPU Training Guide"
]
},
{
"objectID": "docs/dataset-formats/conversation.html",
"href": "docs/dataset-formats/conversation.html",
"title": "Conversation",
"section": "",
"text": "IMPORTANT: ShareGPT is deprecated!. Please see chat_template section below.",
"crumbs": [
"Dataset Formats",
"Conversation"
]
},
{
"objectID": "docs/dataset-formats/conversation.html#sharegpt",
"href": "docs/dataset-formats/conversation.html#sharegpt",
"title": "Conversation",
"section": "",
"text": "IMPORTANT: ShareGPT is deprecated!. Please see chat_template section below.",
"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/conversation.html#chat_template",
"href": "docs/dataset-formats/conversation.html#chat_template",
"title": "Conversation",
"section": "chat_template",
"text": "chat_template\nChat 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\nMigrating from sharegpt\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_field_role: from\n message_field_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_field_role: from\n message_field_content: value\nWe recommend checking the below examples for other usecases.\n\n\nExamples\n\nUsing 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\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(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_field_role: from\n message_field_content: value\n roles_to_train: []\n train_on_eos: turn\n message_field_training: train\n message_field_training_detail: train_detail\nTip: It is not necessary to use both message_field_training and message_field_training_detail at a time.",
"crumbs": [
"Dataset Formats",
"Conversation"
]
},
{
"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\n...",
"crumbs": [
"Dataset Formats",
"Pre-training"
]
},
{
"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/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": [
"How-To Guides",
"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": [
"How-To Guides",
"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] All 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): yaml dataset: ... 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": [
"How-To Guides",
"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] If 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] If 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\": \"${workspaceFolder}/devtools/temp_debug/.hf-cache\"}, // send HF cache to a temp folder\n \"preLaunchTask\": \"cleanup-for-dataprep\", // delete temp folders (see below)\n }\n ]\n}\nAdditional notes about this configuration:\n\nThe argument justMyCode is set to true such that you step through only the axolotl code. If you want to step into dependencies, set this to false.\nThe preLaunchTask: cleanup-for-dataprep is defined in .vscode/tasks.json and is used to delete the following folders before debugging, which is essential to ensure that the data pre-processing code is run from scratch:\n\n./devtools/temp_debug/axolotl_outputs\n./devtools/temp_debug/.hf-cache/datasets\n\n\n\n[!Tip] You may not want to delete these folders. For example, if you are debugging model training instead of data pre-processing, you may NOT want to delete the cache or output folders. You may also need to add additional tasks to the tasks.json file depending on your use case.\n\nBelow is the ./vscode/tasks.json file that defines the cleanup-for-dataprep task. This task is run before each debugging session when you use the above configuration. Note how there are two tasks that delete the two folders mentioned above. The third task cleanup-for-dataprep is a composite task that combines the two tasks. A composite task is necessary because VSCode does not allow you to specify multiple tasks in the preLaunchTask argument of the launch.json file.\n// .vscode/tasks.json\n// this file is used by launch.json\n{\n \"version\": \"2.0.0\",\n \"tasks\": [\n // this task changes into the devtools directory and deletes the temp_debug/axolotl_outputs folder\n {\n \"label\": \"delete-outputs\",\n \"type\": \"shell\",\n \"command\": \"rm -rf temp_debug/axolotl_outputs\",\n \"options\":{ \"cwd\": \"${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\": \"${workspaceFolder}/devtools\"},\n \"problemMatcher\": []\n },\n // this task combines the two tasks above\n {\n \"label\": \"cleanup-for-dataprep\",\n \"dependsOn\": [\"delete-outputs\", \"delete-temp-hf-dataset-cache\"],\n }\n ]\n}\n\n\nCustomizing your debugger\nYour debugging use case may differ from the example above. The easiest thing to do is to put your own axolotl config in the devtools folder and modify the launch.json file to use your config. You may also want to modify the preLaunchTask to delete different folders or not delete anything at all.\n\n\nVideo Tutorial\nThe following video tutorial walks through the above configuration and demonstrates how to debug with VSCode, (click the image below to watch):\n\n\n\nHamel Husain’s tutorial: Debugging Axolotl w/VSCode",
"crumbs": [
"How-To Guides",
"Debugging"
]
},
{
"objectID": "docs/debugging.html#debugging-with-docker",
"href": "docs/debugging.html#debugging-with-docker",
"title": "Debugging",
"section": "Debugging With Docker",
"text": "Debugging With Docker\nUsing official Axolotl Docker images is a great way to debug your code, and is a very popular way to use Axolotl. Attaching VSCode to Docker takes a few more steps.\n\nSetup\nOn the host that is running axolotl (ex: if you are using a remote host), clone the axolotl repo and change your current directory to the root:\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\n\n[!Tip] If you already have axolotl cloned on your host, make sure you have the latest changes and change into the root of the project.\n\nNext, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:2\ndocker run --privileged --gpus '\"all\"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src=\"${PWD}\",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-py3.10-cu118-2.0.1\n\n[!Tip] To 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": [
"How-To Guides",
"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": [
"How-To Guides",
"Debugging"
]
},
{
"objectID": "docs/cli.html",
"href": "docs/cli.html",
"title": "Axolotl CLI Documentation",
"section": "",
"text": "Axolotl CLI Documentation\nThe Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers the CLI commands, their usage, and common examples.\n\nTable of Contents\n\nBasic Commands\nCommand Reference\n\nfetch\npreprocess\ntrain\ninference\nmerge-lora\nmerge-sharded-fsdp-weights\nevaluate\nlm-eval\n\nLegacy CLI Usage\nRemote Compute with Modal Cloud\n\nCloud Configuration\nRunning on Modal Cloud\nCloud Configuration Options\n\n\n\n\nBasic 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.\n\n\nCommand 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\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 using metrics specified in the config.\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\n\n# Evaluate specific tasks\naxolotl lm-eval config.yml --tasks arc_challenge,hellaswag\nConfiguration options:\nlm_eval_tasks: List of tasks to evaluate\nlm_eval_batch_size: Batch size for evaluation\noutput_dir: Directory to save evaluation results\n\n\n\nLegacy 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\nRemote Compute with Modal Cloud\nAxolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a cloud 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\nenv: # Environment variables\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\nenv: Environment variables to pass\nsecrets: Secrets to inject"
},
{
"objectID": "docs/ray-integration.html",
"href": "docs/ray-integration.html",
"title": "Ray Train integration",
"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": [
"How-To Guides",
"Ray Train integration"
]
},
{
"objectID": "docs/ray-integration.html#ray-cluster-setup",
"href": "docs/ray-integration.html#ray-cluster-setup",
"title": "Ray Train integration",
"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: https://docs.ray.io/en/latest/cluster/getting-started.html\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": [
"How-To Guides",
"Ray Train integration"
]
},
{
"objectID": "docs/ray-integration.html#sanity-check",
"href": "docs/ray-integration.html#sanity-check",
"title": "Ray Train integration",
"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": [
"How-To Guides",
"Ray Train integration"
]
},
{
"objectID": "docs/ray-integration.html#configuring-training-with-ray-train",
"href": "docs/ray-integration.html#configuring-training-with-ray-train",
"title": "Ray Train integration",
"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:\n...\nuse_ray: true\nray_num_workers: 4\n# optional\nresources_per_worker:\n GPU: 1\n...\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": [
"How-To Guides",
"Ray Train integration"
]
},
{
"objectID": "docs/ray-integration.html#launching-training",
"href": "docs/ray-integration.html#launching-training",
"title": "Ray Train integration",
"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": [
"How-To Guides",
"Ray Train integration"
]
},
{
"objectID": "docs/torchao.html",
"href": "docs/torchao.html",
"title": "PyTorch ao",
"section": "",
"text": "Installation\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"
},
{
"objectID": "docs/installation.html",
"href": "docs/installation.html",
"title": "Installation Guide",
"section": "",
"text": "This guide covers all the ways you can install and set up Axolotl for your environment.",
"crumbs": [
"How-To Guides",
"Installation Guide"
]
},
{
"objectID": "docs/installation.html#sec-requirements",
"href": "docs/installation.html#sec-requirements",
"title": "Installation Guide",
"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 ≥2.4.1",
"crumbs": [
"How-To Guides",
"Installation Guide"
]
},
{
"objectID": "docs/installation.html#sec-installation-methods",
"href": "docs/installation.html#sec-installation-methods",
"title": "Installation Guide",
"section": "2 Installation Methods",
"text": "2 Installation Methods\n\n2.1 PyPI Installation (Recommended)\npip3 install --no-build-isolation axolotl[flash-attn,deepspeed]\nWe use --no-build-isolation in order to detect the installed PyTorch version (if installed) in order not to clobber it, and so that we set the correct version of dependencies that are specific to the PyTorch version or other installed co-dependencies.\n\n\n2.2 Edge/Development Build\nFor the latest features between releases:\ngit clone https://github.com/axolotl-ai-cloud/axolotl.git\ncd axolotl\npip3 install packaging ninja\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n\n\n2.3 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",
"crumbs": [
"How-To Guides",
"Installation Guide"
]
},
{
"objectID": "docs/installation.html#sec-cloud",
"href": "docs/installation.html#sec-cloud",
"title": "Installation Guide",
"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\n\n\n\n\n3.2 Google Colab\nUse our example notebook.",
"crumbs": [
"How-To Guides",
"Installation Guide"
]
},
{
"objectID": "docs/installation.html#sec-platform-specific",
"href": "docs/installation.html#sec-platform-specific",
"title": "Installation Guide",
"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": [
"How-To Guides",
"Installation Guide"
]
},
{
"objectID": "docs/installation.html#sec-env-managers",
"href": "docs/installation.html#sec-env-managers",
"title": "Installation Guide",
"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 packaging\npip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'\n(Optional) Login to Hugging Face:\nhuggingface-cli login",
"crumbs": [
"How-To Guides",
"Installation Guide"
]
},
{
"objectID": "docs/installation.html#sec-troubleshooting",
"href": "docs/installation.html#sec-troubleshooting",
"title": "Installation Guide",
"section": "6 Troubleshooting",
"text": "6 Troubleshooting\nIf you encounter installation issues, see our FAQ and Debugging Guide.",
"crumbs": [
"How-To Guides",
"Installation Guide"
]
},
{
"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 any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms and conditions set forth in this Agreement.\n\nDefinitions 1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement. 1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl, which may be licensed separately by their respective authors and/or licensors. 1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which permits Plugin Integrations to integrate with the Axolotl service.\nGrant of License 2.1 Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge, publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions: - Licensee must comply with all the terms and conditions of this Agreement. - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial portions of the Software. 2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.\nRestrictions 3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for free or for sale any services, platform, or equivalent to third parties for the purposes of allowing such third parties to fine-tune artificial intelligence models. 3.2 Licensee shall not: - Use the Software for any illegal or unauthorized purpose. - Reverse engineer, decompile, or disassemble the Software. - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software. - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the Software or interfere with any third-party use of the Software. 3.3 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 4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to Licensee.\nDisclaimer of Warranty 5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\nTermination 6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any copies in its possession.\nGoverning Law 7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California, without regards to conflicts of laws provisions thereof.\nEntire Agreement 8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms on a go-forward basis. Axolotl will use commercially reasonable efforts to provide Licensee notice of any material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be bound by the terms and conditions of this Agreement.\n\nThis Agreement was last updated on August 23, 2024."
},
{
"objectID": "src/axolotl/integrations/kd/topk_logprob/LICENSE.html",
"href": "src/axolotl/integrations/kd/topk_logprob/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 any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms and conditions set forth in this Agreement.\n\nDefinitions 1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement. 1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl, which may be licensed separately by their respective authors and/or licensors. 1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which permits Plugin Integrations to integrate with the Axolotl service.\nGrant of License 2.1 Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge, publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions: - Licensee must comply with all the terms and conditions of this Agreement. - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial portions of the Software. 2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.\nRestrictions 3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for free or for sale any services, platform, or equivalent to third parties for the purposes of allowing such third parties to fine-tune artificial intelligence models. 3.2 Licensee shall not: - Use the Software for any illegal or unauthorized purpose. - Reverse engineer, decompile, or disassemble the Software. - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software. - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the Software or interfere with any third-party use of the Software. 3.3 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 4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to Licensee.\nDisclaimer of Warranty 5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.\nTermination 6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any copies in its possession.\nGoverning Law 7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California, without regards to conflicts of laws provisions thereof.\nEntire Agreement 8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms on a go-forward basis. Axolotl will use commercially reasonable efforts to provide Licensee notice of any material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be bound by the terms and conditions of this Agreement.\n\nThis Agreement was last updated on August 23, 2024."
},
{
"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": "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 /arrow/cpp/src/arrow/filesystem/s3fs.cc:2598: arrow::fs::FinalizeS3 was not called even though S3 was initialized. This could lead to a segmentation fault at exit. Try reinstalling bitsandbytes and transformers from source."
},
{
"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 to 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. trainer_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": "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 material, 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": "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 the (dataset format)[../dataset-formats/] and prompt strategies to: - parse the dataset based on the dataset format - transform the dataset to how you would interact with the model based on the prompt strategy - tokenize the dataset based on the configured model & tokenizer - shuffle and merge multiple datasets together if using more than one\nThe processing of the datasets can happen one of two ways:\n\nBefore kicking off training by calling python -m axolotl.cli.preprocess /path/to/your.yaml --debug\nWhen training is started\n\nWhat are the benefits of pre-processing? When training interactively or for sweeps (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent training 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 YAMLs 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 default path of ./last_run_prepared/, but will ignore anything already cached there. By explicitly setting dataset_prepared_path: ./last_run_prepared, the trainer will use whatever pre-processed data is in the cache.\nWhat are the edge cases? Let’s say you are writing a custom prompt strategy or using a user-defined prompt template. Because the trainer cannot readily detect these changes, we cannot change the calculated hash value for the pre-processed dataset. If you have dataset_prepared_path: ... set and change your prompt templating logic, it may not pick up the changes you made and you will be training over the old prompt."
},
{
"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: - FSDP",
"crumbs": [
"How-To Guides",
"Mac M-series"
]
},
{
"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 modules in a model."
},
{
"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 modules in a model."
},
{
"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 of 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 self attention q_proj module."
},
{
"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": [
"How-To Guides",
"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": [
"How-To Guides",
"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] See 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": [
"How-To Guides",
"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": [
"How-To Guides",
"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": [
"How-To Guides",
"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": [
"How-To Guides",
"FDSP + QLoRA"
]
},
{
"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. We 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\n\n\nProcess Reward Models (PRM)\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"
},
{
"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 datasets where each example contains multiple completion steps and a preference label for 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 datasets where each example contains multiple completion steps and a preference label for 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/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/template_free.html",
"href": "docs/dataset-formats/template_free.html",
"title": "Template-Free",
"section": "",
"text": "See these docs.",
"crumbs": [
"Dataset Formats",
"Template-Free"
]
},
{
"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\n\n\n\n\n\nImportant\n\n\n\nFor pre-training only, Axolotl would split texts if it exceeds the context length into multiple smaller prompts.\n\n\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 (either example works):\ndatasets:\n - path: A.jsonl\n type: completion\n\n - path: json\n data_files: [\"A.jsonl\", \"B.jsonl\", \"C.jsonl\"]\n type: completion\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.",
"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 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: - input_ids: from tokenizing formatted prompt - attention_mask: for masking padding. If you don’t add padding, it would be equal to len(input_ids) * [1] - labels: 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\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\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\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\n message_field_role: role\n message_field_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\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. The final step would be to correctly set the EOS token in your config:\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_field_role: role\n message_field_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\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 here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) 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.",
"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 datasets documentation for more detail.",
"crumbs": [
"Dataset Formats"
]
},
{
"objectID": "docs/input_output.html",
"href": "docs/input_output.html",
"title": "Template-free prompt construction",
"section": "",
"text": "Background\n\nMasking Inputs\nYou may not want prompt templates\nThe input_output format\n\nUsage\n\n1. Prepare Data\n2. Use type: input_output\n3. Check the prompts",
"crumbs": [
"How-To Guides",
"Template-free prompt construction"
]
},
{
"objectID": "docs/input_output.html#background",
"href": "docs/input_output.html#background",
"title": "Template-free prompt construction",
"section": "Background",
"text": "Background\n\n\nMasking Inputs\nOne of the most popular features of axolotl is setting the following configuration value:\ntrain_on_inputs: false\nIf you declare a dataset formats such as alpaca or chatml, axolotl knows what is an input (i.e. human) vs. an output (i.e. the assistant) and masks the input labels so that your model can focus on predicting the outputs only.\n\n\n\nYou may not want prompt templates\nHowever, there are many situations where you don’t want to use one of these formats or templates. This is because they can:\n\nAdd unnecessary boilerplate to your prompts.\nCreate artifacts like special delimiters <|im_start|> that can quickly become footguns if you don’t include them correctly at inference time.\nEnforce a chat interface when you do not want one. Sometimes you just want to fine-tune a model to a very specific task and do NOT want multi-turn conversations, roles, etc.\nLimit you to only certain roles that the template allows.\n\n\n\n\nThe input_output format\nYou can construct your prompts without a template by using the input_output format, by setting type: input_output in your configuration 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, type: input_output allows you to mask segments of your text. More details on how this works are described below.",
"crumbs": [
"How-To Guides",
"Template-free prompt construction"
]
},
{
"objectID": "docs/input_output.html#usage",
"href": "docs/input_output.html#usage",
"title": "Template-free prompt construction",
"section": "Usage",
"text": "Usage\nThis is how you can use the input_output format:\n\n\n1. Prepare Data\nTo use the input_output format, collect your data in the following format into a jsonl file (below is the first row from the file output.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 model isn’t trained on it. Some things to keep in mind:\n\n[!IMPORTANT] 1. EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl concatenates all the segments as-is. The tokenizer doesn’t add anything additional. Notice how I added spaces, newlines, <s> (BOS), and </s> (EOS) myself. 2. Make sure you check the materialized output to validate that the prompt is getting assembled how you like.\n\n\n\n\n2. Use type: input_output\nLet’s materialize data with our output.jsonl file by setting type: 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 --debug flag will print the tokens, along with the labels so you can verify that the correct items are being ignored:\n$ python -m axolotl.cli.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, <s>(1, 1) means that the token is <s>, the label is 1 and the token_id is 1. When the label is -100 then that token is ignored for training.\n\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 to 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 version 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": [
"How-To Guides",
"Template-free prompt construction"
]
},
{
"objectID": "docs/multimodal.html",
"href": "docs/multimodal.html",
"title": "MultiModal / Vision Language Models (BETA)",
"section": "",
"text": "MultiModal / Vision Language Models (BETA)\n\nSupported Models\n\nMllama, i.e. llama with vision models\n\n\n\nUsage\nCurrently multimodal support is limited and doesn’t have full feature parity. To finetune a multimodal Llama w/ LoRA, you’ll need to use the following in YAML in combination with the rest of the required hyperparams.\nbase_model: alpindale/Llama-3.2-11B-Vision-Instruct\nprocessor_type: AutoProcessor\nskip_prepare_dataset: true\n\nchat_template: llama3_2_vision\ndatasets:\n - path: HuggingFaceH4/llava-instruct-mix-vsft\n type: chat_template\n split: train[:1%]\n field_messages: messages\nremove_unused_columns: false\nsample_packing: false\n\n# only finetune the Language model, leave the vision model and vision tower frozen\nlora_target_modules: 'language_model.model.layers.[\\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'"
},
{
"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 standard industry baselines.\n\n\nInstallation\nThe following will install the correct unsloth and extras from source.\npython scripts/unsloth_install.py | sh\n\n\nUsing unsloth w Axolotl\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: - 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": [
"How-To Guides",
"Unsloth"
]
},
{
"objectID": "docs/inference.html",
"href": "docs/inference.html",
"title": "Inference Guide",
"section": "",
"text": "This guide covers how to use your trained models for inference, including model loading, interactive testing, and common troubleshooting steps.",
"crumbs": [
"How-To Guides",
"Inference Guide"
]
},
{
"objectID": "docs/inference.html#sec-quickstart",
"href": "docs/inference.html#sec-quickstart",
"title": "Inference Guide",
"section": "1 Quick Start",
"text": "1 Quick Start\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": [
"How-To Guides",
"Inference Guide"
]
},
{
"objectID": "docs/inference.html#sec-advanced",
"href": "docs/inference.html#sec-advanced",
"title": "Inference Guide",
"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": [
"How-To Guides",
"Inference Guide"
]
},
{
"objectID": "docs/inference.html#sec-merging",
"href": "docs/inference.html#sec-merging",
"title": "Inference Guide",
"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": [
"How-To Guides",
"Inference Guide"
]
},
{
"objectID": "docs/inference.html#sec-tokenization",
"href": "docs/inference.html#sec-tokenization",
"title": "Inference Guide",
"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": [
"How-To Guides",
"Inference Guide"
]
},
{
"objectID": "docs/inference.html#sec-troubleshooting",
"href": "docs/inference.html#sec-troubleshooting",
"title": "Inference Guide",
"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": [
"How-To Guides",
"Inference Guide"
]
},
{
"objectID": "docs/faq.html",
"href": "docs/faq.html",
"title": "FAQ",
"section": "",
"text": "Q: 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’\n\nA: You may be using deepspeed with single gpu. Please don’t set deepspeed: in yaml or cli.\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.",
"crumbs": [
"FAQ"
]
},
{
"objectID": "docs/amd_hpc.html",
"href": "docs/amd_hpc.html",
"title": "Training with 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": [
"How-To Guides",
"Training with AMD GPUs on HPC Systems"
]
},
{
"objectID": "docs/amd_hpc.html#setup",
"href": "docs/amd_hpc.html#setup",
"title": "Training with 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: - Edit $HOME/packages/axolotl/src/axolotl/monkeypatch/llama_attn_hijack_flash.py modifying the code to always return False for SwiGLU availability from xformers. - 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": [
"How-To Guides",
"Training with AMD GPUs on HPC Systems"
]
},
{
"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 feedback. Various methods include, but not limited to:\n\nProximal Policy Optimization (PPO) (not yet supported in axolotl)\nDirect Preference Optimization (DPO)\nIdentity Preference Optimization (IPO)\nKahneman-Tversky Optimization (KTO)\nOdds Ratio Preference Optimization (ORPO)",
"crumbs": [
"How-To Guides",
"RLHF (Beta)"
]
},
{
"objectID": "docs/rlhf.html#dpo",
"href": "docs/rlhf.html#dpo",
"title": "RLHF (Beta)",
"section": "DPO",
"text": "DPO\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_field_role: \"role\"\n message_field_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}",
"crumbs": [
"How-To Guides",
"RLHF (Beta)"
]
},
{
"objectID": "docs/rlhf.html#ipo",
"href": "docs/rlhf.html#ipo",
"title": "RLHF (Beta)",
"section": "IPO",
"text": "IPO\nAs IPO is just DPO with a different loss function, all supported options for DPO works here.\nrl: ipo",
"crumbs": [
"How-To Guides",
"RLHF (Beta)"
]
},
{
"objectID": "docs/rlhf.html#orpo",
"href": "docs/rlhf.html#orpo",
"title": "RLHF (Beta)",
"section": "ORPO",
"text": "ORPO\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}",
"crumbs": [
"How-To Guides",
"RLHF (Beta)"
]
},
{
"objectID": "docs/rlhf.html#kto",
"href": "docs/rlhf.html#kto",
"title": "RLHF (Beta)",
"section": "KTO",
"text": "KTO\nrl: kto\nrl_beta: 0.5\nkto_desirable_weight: 0.2\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}",
"crumbs": [
"How-To Guides",
"RLHF (Beta)"
]
},
{
"objectID": "docs/rlhf.html#using-local-dataset-files",
"href": "docs/rlhf.html#using-local-dataset-files",
"title": "RLHF (Beta)",
"section": "Using local dataset files",
"text": "Using local dataset files\ndatasets:\n - ds_type: json\n data_files:\n - orca_rlhf.jsonl\n split: train\n type: chatml.intel",
"crumbs": [
"How-To Guides",
"RLHF (Beta)"
]
},
{
"objectID": "docs/rlhf.html#trl-auto-unwrapping-for-peft",
"href": "docs/rlhf.html#trl-auto-unwrapping-for-peft",
"title": "RLHF (Beta)",
"section": "TRL auto-unwrapping for PEFT",
"text": "TRL 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)"
]
}
]
Reference in New Issue View Git Blame Copy Permalink