diff --git a/.nojekyll b/.nojekyll index b94536d6e..87774211a 100644 --- a/.nojekyll +++ b/.nojekyll @@ -1 +1 @@ -a77e0512 \ No newline at end of file +6fae0227 \ No newline at end of file diff --git a/docs/dataset-formats/index.html b/docs/dataset-formats/index.html index 34696b387..5de9c1264 100644 --- a/docs/dataset-formats/index.html +++ b/docs/dataset-formats/index.html @@ -357,7 +357,7 @@ Description - + Pre-training @@ -365,7 +365,7 @@ Description Data format for a pre-training completion task. - + Instruction Tuning @@ -373,7 +373,7 @@ Description Instruction tuning formats for supervised fine-tuning. - + Conversation @@ -381,7 +381,7 @@ Description Conversation format for supervised fine-tuning. - + Template-Free @@ -389,7 +389,7 @@ Description Construct prompts without a template. - + Custom Pre-Tokenized Dataset diff --git a/docs/torchao.html b/docs/torchao.html new file mode 100644 index 000000000..1c1ae2c00 --- /dev/null +++ b/docs/torchao.html @@ -0,0 +1,757 @@ + + + + + + + + + + +PyTorch ao – Axolotl + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+
+ + +
+ +
+ + +
+ + + +
+ +
+
+

PyTorch ao

+
+ +
+
+ Custom data types and layouts for training and inference +
+
+ + +
+ + + + +
+ + + +
+ + +
+

Installation

+

Stable Release from the PyTorch index

+
pip install torchao --extra-index-url https://download.pytorch.org/whl/cu121 # full options are cpu/cu118/cu121/cu124
+

Nightly release

+
pip install --pre torchao-nightly --index-url https://download.pytorch.org/whl/nightly/cu121 # full options are cpu/cu118/cu121/cu124
+ + +
+ +
+ +
+ + + + + \ No newline at end of file diff --git a/search.json b/search.json index ca35de14f..4a255959c 100644 --- a/search.json +++ b/search.json @@ -66,361 +66,127 @@ ] }, { - "objectID": "docs/faq.html", - "href": "docs/faq.html", - "title": "FAQ", + "objectID": "docs/torchao.html", + "href": "docs/torchao.html", + "title": "PyTorch ao", "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.", - "crumbs": [ - "FAQ" - ] + "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/batch_vs_grad.html", - "href": "docs/batch_vs_grad.html", - "title": "Batch size vs Gradient accumulation", + "objectID": "docs/multi-node.html", + "href": "docs/multi-node.html", + "title": "Multi Node", "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/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", + "text": "You 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\nConfigure your model to use FSDP with for example:", "crumbs": [ "How-To Guides", - "Mac M-series" + "Multi Node" ] }, { - "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 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, sharegpt, 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\n # Optional[str] fastchat conversation type, only used with type: sharegpt\n conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n field_human: # Optional[str]. Human key to use for conversation.\n field_model: # Optional[str]. Assistant key to use for conversation.\n # Add additional keys from your dataset as input or output roles\n roles:\n input: # Optional[List[str]]. These will be masked based on train_on_input\n output: # Optional[List[str]].\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# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\n# 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\n# Saves the desired chat template to the tokenizer_config.json for easier inferencing\n# Currently supports chatml and inst (mistral/mixtral)\nchat_template: chatml\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\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\nhf_mlflow_log_artifacts: # set to true to copy each saved checkpoint on each save to mlflow artifact registry\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\ntorch_compile: # 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, integers for every N steps. decimal for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\nsave_strategy: # Set to `\"no\"` to skip checkpoint saves\nsave_steps: # Leave empty to save at each epoch\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]\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# - 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/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 unsloth from source and downgrade xformers as unsloth is incompatible with the most up to date libraries.\npip install --no-deps \"unsloth @ git+https://github.com/unslothai/unsloth.git\"\npip install --no-deps --force-reinstall xformers==0.0.26.post1\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.", + "objectID": "docs/multi-node.html#machine-configuration", + "href": "docs/multi-node.html#machine-configuration", + "title": "Multi Node", + "section": "Machine configuration", + "text": "Machine configuration\nOn each machine you need 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.\nOn the main machine only, make sure the port you set as main_process_port is open in TCP and reachable by other machines.\nAll you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.", "crumbs": [ "How-To Guides", - "Unsloth" + "Multi Node" ] }, { - "objectID": "docs/dataset-formats/index.html", - "href": "docs/dataset-formats/index.html", - "title": "Dataset Formats", + "objectID": "docs/dataset_preprocessing.html", + "href": "docs/dataset_preprocessing.html", + "title": "Dataset Preprocessing", "section": "", - "text": "Axolotl supports a variety of dataset formats. It is recommended to use a JSONL format. The schema of the JSONL depends upon the task and the prompt template you wish to use. Instead of a JSONL, you can also use a HuggingFace dataset with columns for each JSONL field.\nBelow are these various formats organized by task:\n\n\n\n\n\n\n\n\n\nTitle\n\n\nDescription\n\n\n\n\n\n\nPre-training\n\n\nData format for a pre-training completion task.\n\n\n\n\nInstruction Tuning\n\n\nInstruction tuning formats for supervised fine-tuning.\n\n\n\n\nConversation\n\n\nConversation format for supervised fine-tuning.\n\n\n\n\nTemplate-Free\n\n\nConstruct prompts without a template.\n\n\n\n\nCustom Pre-Tokenized Dataset\n\n\nHow to use a custom pre-tokenized dataset.\n\n\n\n\n\nNo matching items", - "crumbs": [ - "Dataset Formats" - ] + "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/dataset-formats/conversation.html", - "href": "docs/dataset-formats/conversation.html", - "title": "Conversation", + "objectID": "docs/nccl.html", + "href": "docs/nccl.html", + "title": "NCCL", "section": "", - "text": "conversations where from is human/gpt. (optional: first row with role system to override default system prompt)\n\n\ndata.jsonl\n\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\n\nNote: type: sharegpt opens special configs: - conversation: enables conversions to many Conversation types. Refer to the ‘name’ here for options. - roles: allows you to specify the roles for input and output. This is useful for datasets with custom roles such as tool etc to support masking. - field_human: specify the key to use instead of human in the conversation. - field_model: specify the key to use instead of gpt in the conversation.\ndatasets:\n path: ...\n type: sharegpt\n\n conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n field_human: # Optional[str]. Human key to use for conversation.\n field_model: # Optional[str]. Assistant key to use for conversation.\n # Add additional keys from your dataset as input or output roles\n roles:\n input: # Optional[List[str]]. These will be masked based on train_on_input\n output: # Optional[List[str]].", - "crumbs": [ - "Dataset Formats", - "Conversation" - ] - }, - { - "objectID": "docs/dataset-formats/conversation.html#sharegpt", - "href": "docs/dataset-formats/conversation.html#sharegpt", - "title": "Conversation", - "section": "", - "text": "conversations where from is human/gpt. (optional: first row with role system to override default system prompt)\n\n\ndata.jsonl\n\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\n\nNote: type: sharegpt opens special configs: - conversation: enables conversions to many Conversation types. Refer to the ‘name’ here for options. - roles: allows you to specify the roles for input and output. This is useful for datasets with custom roles such as tool etc to support masking. - field_human: specify the key to use instead of human in the conversation. - field_model: specify the key to use instead of gpt in the conversation.\ndatasets:\n path: ...\n type: sharegpt\n\n conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n field_human: # Optional[str]. Human key to use for conversation.\n field_model: # Optional[str]. Assistant key to use for conversation.\n # Add additional keys from your dataset as input or output roles\n roles:\n input: # Optional[List[str]]. These will be masked based on train_on_input\n output: # Optional[List[str]].", - "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#sharegpt.load_role", - "href": "docs/dataset-formats/conversation.html#sharegpt.load_role", - "title": "Conversation", - "section": "sharegpt.load_role", - "text": "sharegpt.load_role\nconversations where role is used instead of from\n\n\ndata.jsonl\n\n{\"conversations\": [{\"role\": \"...\", \"value\": \"...\"}]}", - "crumbs": [ - "Dataset Formats", - "Conversation" - ] - }, - { - "objectID": "docs/dataset-formats/conversation.html#sharegpt.load_guanaco", - "href": "docs/dataset-formats/conversation.html#sharegpt.load_guanaco", - "title": "Conversation", - "section": "sharegpt.load_guanaco", - "text": "sharegpt.load_guanaco\nconversations where from is prompter assistant instead of default sharegpt\n\n\ndata.jsonl\n\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}", - "crumbs": [ - "Dataset Formats", - "Conversation" - ] - }, - { - "objectID": "docs/dataset-formats/conversation.html#sharegpt_jokes", - "href": "docs/dataset-formats/conversation.html#sharegpt_jokes", - "title": "Conversation", - "section": "sharegpt_jokes", - "text": "sharegpt_jokes\ncreates a chat where bot is asked to tell a joke, then explain why the joke is funny\n\n\ndata.jsonl\n\n{\"conversations\": [{\"title\": \"...\", \"text\": \"...\", \"explanation\": \"...\"}]}", - "crumbs": [ - "Dataset Formats", - "Conversation" - ] - }, - { - "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/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]]", + "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", - "Multipack (Sample Packing)" + "NCCL" ] }, { - "objectID": "docs/multipack.html#visualization-of-multipack-with-flash-attention", - "href": "docs/multipack.html#visualization-of-multipack-with-flash-attention", - "title": "Multipack (Sample Packing)", + "objectID": "docs/fsdp_qlora.html", + "href": "docs/fsdp_qlora.html", + "title": "FDSP + QLoRA", "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]]", + "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", - "Multipack (Sample Packing)" + "FDSP + QLoRA" ] }, { - "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.", + "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", - "Multipack (Sample Packing)" + "FDSP + QLoRA" ] }, { - "objectID": "FAQS.html", - "href": "FAQS.html", - "title": "FAQs", + "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/rlhf.html", + "href": "docs/rlhf.html", + "title": "RLHF (Beta)", "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": "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": "examples/colab-notebooks/colab-axolotl-example.html", - "href": "examples/colab-notebooks/colab-axolotl-example.html", - "title": "Example notebook for running Axolotl on google colab", - "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)" - }, - { - "objectID": "examples/colab-notebooks/colab-axolotl-example.html#install-axolotl-and-dependencies", - "href": "examples/colab-notebooks/colab-axolotl-example.html#install-axolotl-and-dependencies", - "title": "Example notebook for running Axolotl on google colab", - "section": "Install Axolotl and dependencies", - "text": "Install Axolotl and dependencies\n\n!pip install torch==\"2.1.2\"\n!pip install -e git+https://github.com/axolotl-ai-cloud/axolotl#egg=axolotl\n!pip install flash-attn==\"2.5.0\"\n!pip install deepspeed==\"0.13.1\"!pip install mlflow==\"2.13.0\"" - }, - { - "objectID": "examples/colab-notebooks/colab-axolotl-example.html#create-an-yaml-config-file", - "href": "examples/colab-notebooks/colab-axolotl-example.html#create-an-yaml-config-file", - "title": "Example notebook for running Axolotl on google colab", - "section": "Create an yaml config file", - "text": "Create an yaml config file\n\nimport yaml\n\n# Your YAML string\nyaml_string = \"\"\"\nbase_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\nmodel_type: LlamaForCausalLM\ntokenizer_type: LlamaTokenizer\n\nload_in_8bit: false\nload_in_4bit: true\nstrict: false\n\ndatasets:\n - path: mhenrichsen/alpaca_2k_test\n type: alpaca\ndataset_prepared_path:\nval_set_size: 0.05\noutput_dir: ./outputs/qlora-out\n\nadapter: qlora\nlora_model_dir:\n\nsequence_len: 4096\nsample_packing: true\neval_sample_packing: false\npad_to_sequence_len: true\n\nlora_r: 32\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\nlora_target_linear: true\nlora_fan_in_fan_out:\n\nwandb_project:\nwandb_entity:\nwandb_watch:\nwandb_name:\nwandb_log_model:\n\ngradient_accumulation_steps: 4\nmicro_batch_size: 2\nnum_epochs: 4\noptimizer: paged_adamw_32bit\nlr_scheduler: cosine\nlearning_rate: 0.0002\n\ntrain_on_inputs: false\ngroup_by_length: false\nbf16: auto\nfp16:\ntf32: false\n\ngradient_checkpointing: true\nearly_stopping_patience:\nresume_from_checkpoint:\nlocal_rank:\nlogging_steps: 1\nxformers_attention:\nflash_attention: true\n\nwarmup_steps: 10\nevals_per_epoch: 4\nsaves_per_epoch: 1\ndebug:\ndeepspeed:\nweight_decay: 0.0\nfsdp:\nfsdp_config:\nspecial_tokens:\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)" - }, - { - "objectID": "examples/colab-notebooks/colab-axolotl-example.html#launch-the-training", - "href": "examples/colab-notebooks/colab-axolotl-example.html#launch-the-training", - "title": "Example notebook for running Axolotl on google colab", - "section": "Launch the training", - "text": "Launch the training\n\n# By using the ! the comand will be executed as a bash command\n!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml" - }, - { - "objectID": "examples/colab-notebooks/colab-axolotl-example.html#play-with-inference", - "href": "examples/colab-notebooks/colab-axolotl-example.html#play-with-inference", - "title": "Example notebook for running Axolotl on google colab", - "section": "Play with inference", - "text": "Play with inference\n\n# By using the ! the comand will be executed as a bash command\n!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n --qlora_model_dir=\"./qlora-out\" --gradio" - }, - { - "objectID": "index.html", - "href": "index.html", - "title": "Axolotl", - "section": "", - "text": "Axolotl\n \n Axolotl supports\n Quickstart ⚡\n \n Usage\n \n Advanced Setup\n \n Environment\n Dataset\n Config\n Train\n Inference Playground\n Merge LORA to base\n \n Common Errors 🧰\n \n Tokenization Mismatch b/w Inference & Training\n \n Debugging Axolotl\n Need help? 🙋\n Badge ❤🏷️\n Community Showcase\n Contributing 🤝\n Sponsors 🤝❤", + "text": "Overview\nReinforcement 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)\n\n\n\nRLHF using Axolotl\n\n[!IMPORTANT] This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.\n\nThe various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML\n\nDPO\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.argilla\n\n\nIPO\nrl: ipo\n\n\nORPO\nPaper: https://arxiv.org/abs/2403.07691\nrl: orpo\norpo_alpha: 0.1\nremove_unused_columns: false\n\nchat_template: chatml\ndatasets:\n - path: argilla/ultrafeedback-binarized-preferences-cleaned\n type: chat_template.argilla\n\n\nUsing local dataset files\ndatasets:\n - ds_type: json\n data_files:\n - orca_rlhf.jsonl\n split: train\n type: chatml.intel\n\n\nTrl autounwrap for peft\nTrl supports autounwrapping peft models, so that a ref model does not need to be additionally loaded, leading to less VRAM needed. This is on by default. To turn it off, pass the following config.\n# load ref model when adapter training.\nrl_adapter_ref_model: true", "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#axolotl-supports", - "href": "index.html#axolotl-supports", - "title": "Axolotl", - "section": "Axolotl supports", - "text": "Axolotl supports\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\n\n✅: supported ❌: not supported ❓: untested", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#quickstart", - "href": "index.html#quickstart", - "title": "Axolotl", - "section": "Quickstart ⚡", - "text": "Quickstart ⚡\nGet started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.\nRequirements: Python >=3.10 and Pytorch >=2.1.1.\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\n\npip3 install packaging ninja\npip3 install -e '.[flash-attn,deepspeed]'\n\nUsage\n# preprocess datasets - optional but recommended\nCUDA_VISIBLE_DEVICES=\"\" python -m axolotl.cli.preprocess examples/openllama-3b/lora.yml\n\n# finetune lora\naccelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml\n\n# inference\naccelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \\\n --lora_model_dir=\"./outputs/lora-out\"\n\n# gradio\naccelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \\\n --lora_model_dir=\"./outputs/lora-out\" --gradio\n\n# remote yaml files - the yaml config can be hosted on a public URL\n# Note: the yaml config must directly link to the **raw** yaml\naccelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/examples/openllama-3b/lora.yml", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#advanced-setup", - "href": "index.html#advanced-setup", - "title": "Axolotl", - "section": "Advanced Setup", - "text": "Advanced Setup\n\nEnvironment\n\nDocker\ndocker run --gpus '\"all\"' --rm -it winglian/axolotl:main-latest\nOr run on the current files for development:\ndocker compose up -d\n\n[!Tip] If you want to debug axolotl or prefer to use Docker as your development environment, see the debugging guide’s section on Docker.\n\n\n\nDocker advanced\n\nA more powerful Docker command to run would be this:\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 ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-latest\nIt additionally: * Prevents memory issues when running e.g. deepspeed (e.g. you could hit SIGBUS/signal 7 error) through --ipc and --ulimit args. * Persists the downloaded HF data (models etc.) and your modifications to axolotl code through --mount/-v args. * The --name argument simply makes it easier to refer to the container in vscode (Dev Containers: Attach to Running Container...) or in your terminal. * The --privileged flag gives all capabilities to the container. * The --shm-size 10g argument increases the shared memory size. Use this if you see exitcode: -7 errors using deepspeed.\nMore information on nvidia website\n\n\n\nConda/Pip venv\n\nInstall python >=3.10\nInstall pytorch stable https://pytorch.org/get-started/locally/\nInstall Axolotl along with python dependencies bash pip3 install packaging pip3 install -e '.[flash-attn,deepspeed]'\n(Optional) Login to Huggingface to use gated models/datasets. bash huggingface-cli login Get the token at huggingface.co/settings/tokens\n\n\n\nCloud GPU\nFor cloud GPU providers that support docker images, use winglian/axolotl-cloud:main-latest\n\non Latitude.sh use this direct link\non JarvisLabs.ai use this direct link\non RunPod use this direct link\n\n\n\nBare Metal Cloud GPU\n\nLambdaLabs\n\n\nClick to Expand\n\n\nInstall python\n\nsudo apt update\nsudo apt install -y python3.10\n\nsudo update-alternatives --install /usr/bin/python python /usr/bin/python3.10 1\nsudo update-alternatives --config python # pick 3.10 if given option\npython -V # should be 3.10\n\nInstall pip\n\nwget https://bootstrap.pypa.io/get-pip.py\npython get-pip.py\n\nInstall Pytorch https://pytorch.org/get-started/locally/\nFollow instructions on quickstart.\nRun\n\npip3 install protobuf==3.20.3\npip3 install -U --ignore-installed requests Pillow psutil scipy\n\nSet path\n\nexport LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH\n\n\n\nGCP\n\n\nClick to Expand\n\nUse a Deeplearning linux OS with cuda and pytorch installed. Then follow instructions on quickstart.\nMake sure to run the below to uninstall xla.\npip uninstall -y torch_xla[tpu]\n\n\n\n\nWindows\nPlease use WSL or Docker!\n\n\nMac\nUse the below instead of the install method in QuickStart.\npip3 install -e '.'\nMore info: mac.md\n\n\nGoogle Colab\nPlease use this example notebook.\n\n\nLaunching on public clouds via SkyPilot\nTo launch on GPU instances (both on-demand and spot instances) on 7+ clouds (GCP, AWS, Azure, OCI, and more), you can use SkyPilot:\npip install \"skypilot-nightly[gcp,aws,azure,oci,lambda,kubernetes,ibm,scp]\" # choose your clouds\nsky check\nGet the example YAMLs of using Axolotl to finetune mistralai/Mistral-7B-v0.1:\ngit clone https://github.com/skypilot-org/skypilot.git\ncd skypilot/llm/axolotl\nUse one command to launch:\n# On-demand\nHF_TOKEN=xx sky launch axolotl.yaml --env HF_TOKEN\n\n# Managed spot (auto-recovery on preemption)\nHF_TOKEN=xx BUCKET=<unique-name> sky spot launch axolotl-spot.yaml --env HF_TOKEN --env BUCKET\n\n\nLaunching on public clouds via dstack\nTo launch on GPU instance (both on-demand and spot instances) on public clouds (GCP, AWS, Azure, Lambda Labs, TensorDock, Vast.ai, and CUDO), you can use dstack.\nWrite a job description in YAML as below:\n# dstack.yaml\ntype: task\n\nimage: winglian/axolotl-cloud:main-20240429-py3.11-cu121-2.2.2\n\nenv:\n - HUGGING_FACE_HUB_TOKEN\n - WANDB_API_KEY\n\ncommands:\n - accelerate launch -m axolotl.cli.train config.yaml\n\nports:\n - 6006\n\nresources:\n gpu:\n memory: 24GB..\n count: 2\nthen, simply run the job with dstack run command. Append --spot option if you want spot instance. dstack run command will show you the instance with cheapest price across multi cloud services:\npip install dstack\nHUGGING_FACE_HUB_TOKEN=xxx WANDB_API_KEY=xxx dstack run . -f dstack.yaml # --spot\nFor further and fine-grained use cases, please refer to the official dstack documents and the detailed description of axolotl example on the official repository.\n\n\n\nDataset\nAxolotl supports a variety of dataset formats. It is recommended to use a JSONL. The schema of the JSONL depends upon the task and the prompt template you wish to use. Instead of a JSONL, you can also use a HuggingFace dataset with columns for each JSONL field.\nSee these docs for more information on how to use different dataset formats.\n\n\nConfig\nSee examples for quick start. It is recommended to duplicate and modify to your needs. The most important options are:\n\nmodel\nbase_model: ./llama-7b-hf # local or huggingface repo\nNote: The code will load the right architecture.\ndataset\ndatasets:\n # huggingface repo\n - path: vicgalle/alpaca-gpt4\n type: alpaca\n\n # huggingface repo with specific configuration/subset\n - path: EleutherAI/pile\n name: enron_emails\n type: completion # format from earlier\n field: text # Optional[str] default: text, field to use for completion data\n\n # huggingface repo with multiple named configurations/subsets\n - path: bigcode/commitpackft\n name:\n - ruby\n - python\n - typescript\n type: ... # unimplemented custom format\n\n # fastchat conversation\n # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n - path: ...\n type: sharegpt\n conversation: chatml # default: vicuna_v1.1\n\n # local\n - path: data.jsonl # or json\n ds_type: json # see other options below\n type: alpaca\n\n # dataset with splits, but no train split\n - path: knowrohit07/know_sql\n type: context_qa.load_v2\n train_on_split: validation\n\n # loading from s3 or gcs\n # s3 creds will be loaded from the system default and gcs only supports public access\n - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above. Supports s3, gcs.\n ...\n\n # Loading Data From a Public URL\n # - The file format is `json` (which includes `jsonl`) by default. For different formats, adjust the `ds_type` option accordingly.\n - path: https://some.url.com/yourdata.jsonl # The URL should be a direct link to the file you wish to load. URLs must use HTTPS protocol, not HTTP.\n ds_type: json # this is the default, see other options below.\nloading\nload_in_4bit: true\nload_in_8bit: true\n\nbf16: auto # require >=ampere, auto will detect if your GPU supports this and choose automatically.\nfp16: # leave empty to use fp16 when bf16 is 'auto'. set to false if you want to fallback to fp32\ntf32: true # require >=ampere\n\nbfloat16: true # require >=ampere, use instead of bf16 when you don't want AMP (automatic mixed precision)\nfloat16: true # use instead of fp16 when you don't want AMP\nNote: Repo does not do 4-bit quantization.\nlora\nadapter: lora # 'qlora' or leave blank for full finetune\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n - q_proj\n - v_proj\n\n\nAll Config Options\nSee these docs for all config options.\n\n\n\nTrain\nRun\naccelerate launch -m axolotl.cli.train your_config.yml\n\n[!TIP] You can also reference a config file that is hosted on a public URL, for example accelerate launch -m axolotl.cli.train https://yourdomain.com/your_config.yml\n\n\nPreprocess dataset\nYou can optionally pre-tokenize dataset with the following before finetuning. This is recommended for large datasets.\n\nSet dataset_prepared_path: to a local folder for saving and loading pre-tokenized dataset.\n(Optional): Set push_dataset_to_hub: hf_user/repo to push it to Huggingface.\n(Optional): Use --debug to see preprocessed examples.\n\npython -m axolotl.cli.preprocess your_config.yml\n\n\nMulti-GPU\nBelow are the options available in axolotl for training with multiple GPUs. Note that DeepSpeed is the recommended multi-GPU option currently because FSDP may experience loss instability.\n\nDeepSpeed\nDeepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you might typically be able to fit into your GPU’s VRAM. More information about the various optimization types for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated\nWe provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.\ndeepspeed: deepspeed_configs/zero1.json\naccelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json\n\n\nFSDP\n\nllama FSDP\n\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\nFSDP + QLoRA\nAxolotl supports training with FSDP and QLoRA, see these docs for more information.\n\n\nWeights & Biases Logging\nMake sure your WANDB_API_KEY environment variable is set (recommended) or you login to wandb with wandb login.\n\nwandb options\n\nwandb_mode:\nwandb_project:\nwandb_entity:\nwandb_watch:\nwandb_name:\nwandb_log_model:\n\n\nSpecial Tokens\nIt is important to have special tokens like delimiters, end-of-sequence, beginning-of-sequence in your tokenizer’s vocabulary. This will help you avoid tokenization issues and help your model train better. You can do this in axolotl like this:\nspecial_tokens:\n bos_token: \"<s>\"\n eos_token: \"</s>\"\n unk_token: \"<unk>\"\ntokens: # these are delimiters\n - \"<|im_start|>\"\n - \"<|im_end|>\"\nWhen you include these tokens in your axolotl config, axolotl adds these tokens to the tokenizer’s vocabulary.\n\n\n\n\nInference Playground\nAxolotl allows you to load your model in an interactive terminal playground for quick experimentation. The config file is the same config file used for training.\nPass the appropriate flag to the inference command, depending upon what kind of model was trained:\n\nPretrained LORA:\npython -m axolotl.cli.inference examples/your_config.yml --lora_model_dir=\"./lora-output-dir\"\nFull weights finetune:\npython -m axolotl.cli.inference examples/your_config.yml --base_model=\"./completed-model\"\nFull weights finetune w/ a prompt from a text file:\ncat /tmp/prompt.txt | python -m axolotl.cli.inference examples/your_config.yml \\\n --base_model=\"./completed-model\" --prompter=None --load_in_8bit=True\n– With gradio hosting\npython -m axolotl.cli.inference examples/your_config.yml --gradio\n\nPlease use --sample_packing False if you have it on and receive the error similar to below:\n\nRuntimeError: stack expects each tensor to be equal size, but got [1, 32, 1, 128] at entry 0 and [1, 32, 8, 128] at entry 1\n\n\n\nMerge LORA to base\nThe following command will merge your LORA adapater with your base model. You can optionally pass the argument --lora_model_dir to specify the directory where your LORA adapter was saved, otherwhise, this will be inferred from output_dir in your axolotl config file. The merged model is saved in the sub-directory {lora_model_dir}/merged.\npython3 -m axolotl.cli.merge_lora your_config.yml --lora_model_dir=\"./completed-model\"\nYou may need to use the gpu_memory_limit and/or lora_on_cpu config options to avoid running out of memory. If you still run out of CUDA memory, you can try to merge in system RAM with\nCUDA_VISIBLE_DEVICES=\"\" python3 -m axolotl.cli.merge_lora ...\nalthough this will be very slow, and using the config options above are recommended instead.", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#common-errors", - "href": "index.html#common-errors", - "title": "Axolotl", - "section": "Common Errors 🧰", - "text": "Common Errors 🧰\nSee also the FAQ’s and debugging guide.\n\nIf you encounter a ‘Cuda out of memory’ error, it means your GPU ran out of memory during the training process. Here’s how to resolve it:\n\nPlease reduce any below - micro_batch_size - eval_batch_size - gradient_accumulation_steps - sequence_len\nIf it does not help, try running without deepspeed and without accelerate (replace “accelerate launch” with “python”) in the command.\nUsing adamw_bnb_8bit might also save you some memory.\n\nfailed (exitcode: -9)\n\nUsually means your system has run out of system memory. Similarly, you should consider reducing the same settings as when you run out of VRAM. Additionally, look into upgrading your system RAM which should be simpler than GPU upgrades.\n\nRuntimeError: expected scalar type Float but found Half\n\nTry set fp16: true\n\nNotImplementedError: No operator found for memory_efficient_attention_forward …\n\nTry to turn off xformers.\n\naccelerate config missing\n\nIt’s safe to ignore it.\n\nNCCL Timeouts during training\n\nSee the NCCL guide.\n\nTokenization Mismatch b/w Inference & Training\nFor many formats, Axolotl constructs prompts by concatenating token ids after tokenizing strings. The reason for concatenating token ids rather than operating on strings is to maintain precise accounting for attention masks.\nIf you decode a prompt constructed by axolotl, you might see spaces between tokens (or lack thereof) that you do not expect, especially around delimiters and special tokens. When you are starting out with a new format, you should always do the following:\n\nMaterialize some data using python -m axolotl.cli.preprocess your_config.yml --debug, and then decode the first few rows with your model’s tokenizer.\nDuring inference, right before you pass a tensor of token ids to your model, decode these tokens back into a string.\nMake sure the inference string from #2 looks exactly like the data you fine tuned on from #1, including spaces and new lines. If they aren’t the same, adjust your inference server accordingly.\nAs an additional troubleshooting step, you can look at the token ids between 1 and 2 to make sure they are identical.\n\nHaving misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this. See this blog post for a concrete example.", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#debugging-axolotl", - "href": "index.html#debugging-axolotl", - "title": "Axolotl", - "section": "Debugging Axolotl", - "text": "Debugging Axolotl\nSee this debugging guide for tips on debugging Axolotl, along with an example configuration for debugging with VSCode.", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#need-help", - "href": "index.html#need-help", - "title": "Axolotl", - "section": "Need help? 🙋", - "text": "Need help? 🙋\nJoin our Discord server where we our community members can help you.\nNeed dedicated support? Please contact us at ✉️wing@openaccessaicollective.org for dedicated support options.", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#badge", - "href": "index.html#badge", - "title": "Axolotl", - "section": "Badge ❤🏷️", - "text": "Badge ❤🏷️\nBuilding something cool with Axolotl? Consider adding a badge to your model card.\n[<img src=\"https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png\" alt=\"Built with Axolotl\" width=\"200\" height=\"32\"/>](https://github.com/axolotl-ai-cloud/axolotl)", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#community-showcase", - "href": "index.html#community-showcase", - "title": "Axolotl", - "section": "Community Showcase", - "text": "Community Showcase\nCheck out some of the projects and models that have been built using Axolotl! Have a model you’d like to add to our Community Showcase? Open a PR with your model.\nOpen Access AI Collective - Minotaur 13b - Manticore 13b - Hippogriff 30b\nPocketDoc Labs - Dan’s PersonalityEngine 13b LoRA", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#contributing", - "href": "index.html#contributing", - "title": "Axolotl", - "section": "Contributing 🤝", - "text": "Contributing 🤝\nPlease read the contributing guide\nBugs? Please check the open issues else create a new Issue.\nPRs are greatly welcome!\nPlease run the quickstart instructions followed by the below to setup env:\npip3 install -r requirements-dev.txt -r requirements-tests.txt\npre-commit install\n\n# test\npytest tests/\n\n# optional: run against all files\npre-commit run --all-files\nThanks to all of our contributors to date. Help drive open source AI progress forward by contributing to Axolotl.", - "crumbs": [ - "Home" - ] - }, - { - "objectID": "index.html#sponsors", - "href": "index.html#sponsors", - "title": "Axolotl", - "section": "Sponsors 🤝❤", - "text": "Sponsors 🤝❤\nOpenAccess AI Collective is run by volunteer contributors such as winglian, NanoCode012, tmm1, mhenrichsen, casper-hansen, hamelsmu and many more who help us accelerate forward by fixing bugs, answering community questions and implementing new features. Axolotl needs donations from sponsors for the compute needed to run our unit & integration tests, troubleshooting community issues, and providing bounties. If you love axolotl, consider sponsoring the project via GitHub Sponsors, Ko-fi or reach out directly to wing@openaccessaicollective.org.\n\n\n💎 Diamond Sponsors - Contact directly\n\n\n\n🥇 Gold Sponsors - $5000/mo\n\n\n\n🥈 Silver Sponsors - $1000/mo\n\n\n\n🥉 Bronze Sponsors - $500/mo\n\nJarvisLabs.ai", - "crumbs": [ - "Home" - ] - }, - { - "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: # hf path only\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.\nDo not add BOS/EOS. Axolotl will add them for you based on the default tokenizer for the model you’re using.\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" + "How-To Guides", + "RLHF (Beta)" ] }, { @@ -666,120 +432,361 @@ ] }, { - "objectID": "docs/rlhf.html", - "href": "docs/rlhf.html", - "title": "RLHF (Beta)", + "objectID": "docs/dataset-formats/tokenized.html", + "href": "docs/dataset-formats/tokenized.html", + "title": "Custom Pre-Tokenized Dataset", "section": "", - "text": "Overview\nReinforcement 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)\n\n\n\nRLHF using Axolotl\n\n[!IMPORTANT] This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.\n\nThe various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML\n\nDPO\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.argilla\n\n\nIPO\nrl: ipo\n\n\nORPO\nPaper: https://arxiv.org/abs/2403.07691\nrl: orpo\norpo_alpha: 0.1\nremove_unused_columns: false\n\nchat_template: chatml\ndatasets:\n - path: argilla/ultrafeedback-binarized-preferences-cleaned\n type: chat_template.argilla\n\n\nUsing local dataset files\ndatasets:\n - ds_type: json\n data_files:\n - orca_rlhf.jsonl\n split: train\n type: chatml.intel\n\n\nTrl autounwrap for peft\nTrl supports autounwrapping peft models, so that a ref model does not need to be additionally loaded, leading to less VRAM needed. This is on by default. To turn it off, pass the following config.\n# load ref model when adapter training.\nrl_adapter_ref_model: true", + "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.\nDo not add BOS/EOS. Axolotl will add them for you based on the default tokenizer for the model you’re using.\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": [ - "How-To Guides", - "RLHF (Beta)" + "Dataset Formats", + "Custom Pre-Tokenized Dataset" ] }, { - "objectID": "docs/fsdp_qlora.html", - "href": "docs/fsdp_qlora.html", - "title": "FDSP + QLoRA", + "objectID": "docs/dataset-formats/pretraining.html", + "href": "docs/dataset-formats/pretraining.html", + "title": "Pre-training", "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.", + "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: # hf path only\n...", "crumbs": [ - "How-To Guides", - "FDSP + QLoRA" + "Dataset Formats", + "Pre-training" ] }, { - "objectID": "docs/fsdp_qlora.html#background", - "href": "docs/fsdp_qlora.html#background", - "title": "FDSP + QLoRA", + "objectID": "index.html", + "href": "index.html", + "title": "Axolotl", "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.", + "text": "Axolotl\n \n Axolotl supports\n Quickstart ⚡\n \n Usage\n \n Advanced Setup\n \n Environment\n Dataset\n Config\n Train\n Inference Playground\n Merge LORA to base\n \n Common Errors 🧰\n \n Tokenization Mismatch b/w Inference & Training\n \n Debugging Axolotl\n Need help? 🙋\n Badge ❤🏷️\n Community Showcase\n Contributing 🤝\n Sponsors 🤝❤", "crumbs": [ - "How-To Guides", - "FDSP + QLoRA" + "Home" ] }, { - "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.", + "objectID": "index.html#axolotl-supports", + "href": "index.html#axolotl-supports", + "title": "Axolotl", + "section": "Axolotl supports", + "text": "Axolotl supports\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\n\n✅: supported ❌: not supported ❓: untested", "crumbs": [ - "How-To Guides", - "FDSP + QLoRA" + "Home" ] }, { - "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.", + "objectID": "index.html#quickstart", + "href": "index.html#quickstart", + "title": "Axolotl", + "section": "Quickstart ⚡", + "text": "Quickstart ⚡\nGet started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.\nRequirements: Python >=3.10 and Pytorch >=2.1.1.\ngit clone https://github.com/axolotl-ai-cloud/axolotl\ncd axolotl\n\npip3 install packaging ninja\npip3 install -e '.[flash-attn,deepspeed]'\n\nUsage\n# preprocess datasets - optional but recommended\nCUDA_VISIBLE_DEVICES=\"\" python -m axolotl.cli.preprocess examples/openllama-3b/lora.yml\n\n# finetune lora\naccelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml\n\n# inference\naccelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \\\n --lora_model_dir=\"./outputs/lora-out\"\n\n# gradio\naccelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \\\n --lora_model_dir=\"./outputs/lora-out\" --gradio\n\n# remote yaml files - the yaml config can be hosted on a public URL\n# Note: the yaml config must directly link to the **raw** yaml\naccelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/examples/openllama-3b/lora.yml", "crumbs": [ - "How-To Guides", - "FDSP + QLoRA" + "Home" ] }, { - "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", + "objectID": "index.html#advanced-setup", + "href": "index.html#advanced-setup", + "title": "Axolotl", + "section": "Advanced Setup", + "text": "Advanced Setup\n\nEnvironment\n\nDocker\ndocker run --gpus '\"all\"' --rm -it winglian/axolotl:main-latest\nOr run on the current files for development:\ndocker compose up -d\n\n[!Tip] If you want to debug axolotl or prefer to use Docker as your development environment, see the debugging guide’s section on Docker.\n\n\n\nDocker advanced\n\nA more powerful Docker command to run would be this:\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 ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-latest\nIt additionally: * Prevents memory issues when running e.g. deepspeed (e.g. you could hit SIGBUS/signal 7 error) through --ipc and --ulimit args. * Persists the downloaded HF data (models etc.) and your modifications to axolotl code through --mount/-v args. * The --name argument simply makes it easier to refer to the container in vscode (Dev Containers: Attach to Running Container...) or in your terminal. * The --privileged flag gives all capabilities to the container. * The --shm-size 10g argument increases the shared memory size. Use this if you see exitcode: -7 errors using deepspeed.\nMore information on nvidia website\n\n\n\nConda/Pip venv\n\nInstall python >=3.10\nInstall pytorch stable https://pytorch.org/get-started/locally/\nInstall Axolotl along with python dependencies bash pip3 install packaging pip3 install -e '.[flash-attn,deepspeed]'\n(Optional) Login to Huggingface to use gated models/datasets. bash huggingface-cli login Get the token at huggingface.co/settings/tokens\n\n\n\nCloud GPU\nFor cloud GPU providers that support docker images, use winglian/axolotl-cloud:main-latest\n\non Latitude.sh use this direct link\non JarvisLabs.ai use this direct link\non RunPod use this direct link\n\n\n\nBare Metal Cloud GPU\n\nLambdaLabs\n\n\nClick to Expand\n\n\nInstall python\n\nsudo apt update\nsudo apt install -y python3.10\n\nsudo update-alternatives --install /usr/bin/python python /usr/bin/python3.10 1\nsudo update-alternatives --config python # pick 3.10 if given option\npython -V # should be 3.10\n\nInstall pip\n\nwget https://bootstrap.pypa.io/get-pip.py\npython get-pip.py\n\nInstall Pytorch https://pytorch.org/get-started/locally/\nFollow instructions on quickstart.\nRun\n\npip3 install protobuf==3.20.3\npip3 install -U --ignore-installed requests Pillow psutil scipy\n\nSet path\n\nexport LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH\n\n\n\nGCP\n\n\nClick to Expand\n\nUse a Deeplearning linux OS with cuda and pytorch installed. Then follow instructions on quickstart.\nMake sure to run the below to uninstall xla.\npip uninstall -y torch_xla[tpu]\n\n\n\n\nWindows\nPlease use WSL or Docker!\n\n\nMac\nUse the below instead of the install method in QuickStart.\npip3 install -e '.'\nMore info: mac.md\n\n\nGoogle Colab\nPlease use this example notebook.\n\n\nLaunching on public clouds via SkyPilot\nTo launch on GPU instances (both on-demand and spot instances) on 7+ clouds (GCP, AWS, Azure, OCI, and more), you can use SkyPilot:\npip install \"skypilot-nightly[gcp,aws,azure,oci,lambda,kubernetes,ibm,scp]\" # choose your clouds\nsky check\nGet the example YAMLs of using Axolotl to finetune mistralai/Mistral-7B-v0.1:\ngit clone https://github.com/skypilot-org/skypilot.git\ncd skypilot/llm/axolotl\nUse one command to launch:\n# On-demand\nHF_TOKEN=xx sky launch axolotl.yaml --env HF_TOKEN\n\n# Managed spot (auto-recovery on preemption)\nHF_TOKEN=xx BUCKET=<unique-name> sky spot launch axolotl-spot.yaml --env HF_TOKEN --env BUCKET\n\n\nLaunching on public clouds via dstack\nTo launch on GPU instance (both on-demand and spot instances) on public clouds (GCP, AWS, Azure, Lambda Labs, TensorDock, Vast.ai, and CUDO), you can use dstack.\nWrite a job description in YAML as below:\n# dstack.yaml\ntype: task\n\nimage: winglian/axolotl-cloud:main-20240429-py3.11-cu121-2.2.2\n\nenv:\n - HUGGING_FACE_HUB_TOKEN\n - WANDB_API_KEY\n\ncommands:\n - accelerate launch -m axolotl.cli.train config.yaml\n\nports:\n - 6006\n\nresources:\n gpu:\n memory: 24GB..\n count: 2\nthen, simply run the job with dstack run command. Append --spot option if you want spot instance. dstack run command will show you the instance with cheapest price across multi cloud services:\npip install dstack\nHUGGING_FACE_HUB_TOKEN=xxx WANDB_API_KEY=xxx dstack run . -f dstack.yaml # --spot\nFor further and fine-grained use cases, please refer to the official dstack documents and the detailed description of axolotl example on the official repository.\n\n\n\nDataset\nAxolotl supports a variety of dataset formats. It is recommended to use a JSONL. The schema of the JSONL depends upon the task and the prompt template you wish to use. Instead of a JSONL, you can also use a HuggingFace dataset with columns for each JSONL field.\nSee these docs for more information on how to use different dataset formats.\n\n\nConfig\nSee examples for quick start. It is recommended to duplicate and modify to your needs. The most important options are:\n\nmodel\nbase_model: ./llama-7b-hf # local or huggingface repo\nNote: The code will load the right architecture.\ndataset\ndatasets:\n # huggingface repo\n - path: vicgalle/alpaca-gpt4\n type: alpaca\n\n # huggingface repo with specific configuration/subset\n - path: EleutherAI/pile\n name: enron_emails\n type: completion # format from earlier\n field: text # Optional[str] default: text, field to use for completion data\n\n # huggingface repo with multiple named configurations/subsets\n - path: bigcode/commitpackft\n name:\n - ruby\n - python\n - typescript\n type: ... # unimplemented custom format\n\n # fastchat conversation\n # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n - path: ...\n type: sharegpt\n conversation: chatml # default: vicuna_v1.1\n\n # local\n - path: data.jsonl # or json\n ds_type: json # see other options below\n type: alpaca\n\n # dataset with splits, but no train split\n - path: knowrohit07/know_sql\n type: context_qa.load_v2\n train_on_split: validation\n\n # loading from s3 or gcs\n # s3 creds will be loaded from the system default and gcs only supports public access\n - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above. Supports s3, gcs.\n ...\n\n # Loading Data From a Public URL\n # - The file format is `json` (which includes `jsonl`) by default. For different formats, adjust the `ds_type` option accordingly.\n - path: https://some.url.com/yourdata.jsonl # The URL should be a direct link to the file you wish to load. URLs must use HTTPS protocol, not HTTP.\n ds_type: json # this is the default, see other options below.\nloading\nload_in_4bit: true\nload_in_8bit: true\n\nbf16: auto # require >=ampere, auto will detect if your GPU supports this and choose automatically.\nfp16: # leave empty to use fp16 when bf16 is 'auto'. set to false if you want to fallback to fp32\ntf32: true # require >=ampere\n\nbfloat16: true # require >=ampere, use instead of bf16 when you don't want AMP (automatic mixed precision)\nfloat16: true # use instead of fp16 when you don't want AMP\nNote: Repo does not do 4-bit quantization.\nlora\nadapter: lora # 'qlora' or leave blank for full finetune\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n - q_proj\n - v_proj\n\n\nAll Config Options\nSee these docs for all config options.\n\n\n\nTrain\nRun\naccelerate launch -m axolotl.cli.train your_config.yml\n\n[!TIP] You can also reference a config file that is hosted on a public URL, for example accelerate launch -m axolotl.cli.train https://yourdomain.com/your_config.yml\n\n\nPreprocess dataset\nYou can optionally pre-tokenize dataset with the following before finetuning. This is recommended for large datasets.\n\nSet dataset_prepared_path: to a local folder for saving and loading pre-tokenized dataset.\n(Optional): Set push_dataset_to_hub: hf_user/repo to push it to Huggingface.\n(Optional): Use --debug to see preprocessed examples.\n\npython -m axolotl.cli.preprocess your_config.yml\n\n\nMulti-GPU\nBelow are the options available in axolotl for training with multiple GPUs. Note that DeepSpeed is the recommended multi-GPU option currently because FSDP may experience loss instability.\n\nDeepSpeed\nDeepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you might typically be able to fit into your GPU’s VRAM. More information about the various optimization types for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated\nWe provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.\ndeepspeed: deepspeed_configs/zero1.json\naccelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json\n\n\nFSDP\n\nllama FSDP\n\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\nFSDP + QLoRA\nAxolotl supports training with FSDP and QLoRA, see these docs for more information.\n\n\nWeights & Biases Logging\nMake sure your WANDB_API_KEY environment variable is set (recommended) or you login to wandb with wandb login.\n\nwandb options\n\nwandb_mode:\nwandb_project:\nwandb_entity:\nwandb_watch:\nwandb_name:\nwandb_log_model:\n\n\nSpecial Tokens\nIt is important to have special tokens like delimiters, end-of-sequence, beginning-of-sequence in your tokenizer’s vocabulary. This will help you avoid tokenization issues and help your model train better. You can do this in axolotl like this:\nspecial_tokens:\n bos_token: \"<s>\"\n eos_token: \"</s>\"\n unk_token: \"<unk>\"\ntokens: # these are delimiters\n - \"<|im_start|>\"\n - \"<|im_end|>\"\nWhen you include these tokens in your axolotl config, axolotl adds these tokens to the tokenizer’s vocabulary.\n\n\n\n\nInference Playground\nAxolotl allows you to load your model in an interactive terminal playground for quick experimentation. The config file is the same config file used for training.\nPass the appropriate flag to the inference command, depending upon what kind of model was trained:\n\nPretrained LORA:\npython -m axolotl.cli.inference examples/your_config.yml --lora_model_dir=\"./lora-output-dir\"\nFull weights finetune:\npython -m axolotl.cli.inference examples/your_config.yml --base_model=\"./completed-model\"\nFull weights finetune w/ a prompt from a text file:\ncat /tmp/prompt.txt | python -m axolotl.cli.inference examples/your_config.yml \\\n --base_model=\"./completed-model\" --prompter=None --load_in_8bit=True\n– With gradio hosting\npython -m axolotl.cli.inference examples/your_config.yml --gradio\n\nPlease use --sample_packing False if you have it on and receive the error similar to below:\n\nRuntimeError: stack expects each tensor to be equal size, but got [1, 32, 1, 128] at entry 0 and [1, 32, 8, 128] at entry 1\n\n\n\nMerge LORA to base\nThe following command will merge your LORA adapater with your base model. You can optionally pass the argument --lora_model_dir to specify the directory where your LORA adapter was saved, otherwhise, this will be inferred from output_dir in your axolotl config file. The merged model is saved in the sub-directory {lora_model_dir}/merged.\npython3 -m axolotl.cli.merge_lora your_config.yml --lora_model_dir=\"./completed-model\"\nYou may need to use the gpu_memory_limit and/or lora_on_cpu config options to avoid running out of memory. If you still run out of CUDA memory, you can try to merge in system RAM with\nCUDA_VISIBLE_DEVICES=\"\" python3 -m axolotl.cli.merge_lora ...\nalthough this will be very slow, and using the config options above are recommended instead.", "crumbs": [ - "How-To Guides", - "FDSP + QLoRA" + "Home" ] }, { - "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.↩︎", + "objectID": "index.html#common-errors", + "href": "index.html#common-errors", + "title": "Axolotl", + "section": "Common Errors 🧰", + "text": "Common Errors 🧰\nSee also the FAQ’s and debugging guide.\n\nIf you encounter a ‘Cuda out of memory’ error, it means your GPU ran out of memory during the training process. Here’s how to resolve it:\n\nPlease reduce any below - micro_batch_size - eval_batch_size - gradient_accumulation_steps - sequence_len\nIf it does not help, try running without deepspeed and without accelerate (replace “accelerate launch” with “python”) in the command.\nUsing adamw_bnb_8bit might also save you some memory.\n\nfailed (exitcode: -9)\n\nUsually means your system has run out of system memory. Similarly, you should consider reducing the same settings as when you run out of VRAM. Additionally, look into upgrading your system RAM which should be simpler than GPU upgrades.\n\nRuntimeError: expected scalar type Float but found Half\n\nTry set fp16: true\n\nNotImplementedError: No operator found for memory_efficient_attention_forward …\n\nTry to turn off xformers.\n\naccelerate config missing\n\nIt’s safe to ignore it.\n\nNCCL Timeouts during training\n\nSee the NCCL guide.\n\nTokenization Mismatch b/w Inference & Training\nFor many formats, Axolotl constructs prompts by concatenating token ids after tokenizing strings. The reason for concatenating token ids rather than operating on strings is to maintain precise accounting for attention masks.\nIf you decode a prompt constructed by axolotl, you might see spaces between tokens (or lack thereof) that you do not expect, especially around delimiters and special tokens. When you are starting out with a new format, you should always do the following:\n\nMaterialize some data using python -m axolotl.cli.preprocess your_config.yml --debug, and then decode the first few rows with your model’s tokenizer.\nDuring inference, right before you pass a tensor of token ids to your model, decode these tokens back into a string.\nMake sure the inference string from #2 looks exactly like the data you fine tuned on from #1, including spaces and new lines. If they aren’t the same, adjust your inference server accordingly.\nAs an additional troubleshooting step, you can look at the token ids between 1 and 2 to make sure they are identical.\n\nHaving misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this. See this blog post for a concrete example.", "crumbs": [ - "How-To Guides", - "FDSP + QLoRA" + "Home" ] }, { - "objectID": "docs/nccl.html", - "href": "docs/nccl.html", - "title": "NCCL", + "objectID": "index.html#debugging-axolotl", + "href": "index.html#debugging-axolotl", + "title": "Axolotl", + "section": "Debugging Axolotl", + "text": "Debugging Axolotl\nSee this debugging guide for tips on debugging Axolotl, along with an example configuration for debugging with VSCode.", + "crumbs": [ + "Home" + ] + }, + { + "objectID": "index.html#need-help", + "href": "index.html#need-help", + "title": "Axolotl", + "section": "Need help? 🙋", + "text": "Need help? 🙋\nJoin our Discord server where we our community members can help you.\nNeed dedicated support? Please contact us at ✉️wing@openaccessaicollective.org for dedicated support options.", + "crumbs": [ + "Home" + ] + }, + { + "objectID": "index.html#badge", + "href": "index.html#badge", + "title": "Axolotl", + "section": "Badge ❤🏷️", + "text": "Badge ❤🏷️\nBuilding something cool with Axolotl? Consider adding a badge to your model card.\n[<img src=\"https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png\" alt=\"Built with Axolotl\" width=\"200\" height=\"32\"/>](https://github.com/axolotl-ai-cloud/axolotl)", + "crumbs": [ + "Home" + ] + }, + { + "objectID": "index.html#community-showcase", + "href": "index.html#community-showcase", + "title": "Axolotl", + "section": "Community Showcase", + "text": "Community Showcase\nCheck out some of the projects and models that have been built using Axolotl! Have a model you’d like to add to our Community Showcase? Open a PR with your model.\nOpen Access AI Collective - Minotaur 13b - Manticore 13b - Hippogriff 30b\nPocketDoc Labs - Dan’s PersonalityEngine 13b LoRA", + "crumbs": [ + "Home" + ] + }, + { + "objectID": "index.html#contributing", + "href": "index.html#contributing", + "title": "Axolotl", + "section": "Contributing 🤝", + "text": "Contributing 🤝\nPlease read the contributing guide\nBugs? Please check the open issues else create a new Issue.\nPRs are greatly welcome!\nPlease run the quickstart instructions followed by the below to setup env:\npip3 install -r requirements-dev.txt -r requirements-tests.txt\npre-commit install\n\n# test\npytest tests/\n\n# optional: run against all files\npre-commit run --all-files\nThanks to all of our contributors to date. Help drive open source AI progress forward by contributing to Axolotl.", + "crumbs": [ + "Home" + ] + }, + { + "objectID": "index.html#sponsors", + "href": "index.html#sponsors", + "title": "Axolotl", + "section": "Sponsors 🤝❤", + "text": "Sponsors 🤝❤\nOpenAccess AI Collective is run by volunteer contributors such as winglian, NanoCode012, tmm1, mhenrichsen, casper-hansen, hamelsmu and many more who help us accelerate forward by fixing bugs, answering community questions and implementing new features. Axolotl needs donations from sponsors for the compute needed to run our unit & integration tests, troubleshooting community issues, and providing bounties. If you love axolotl, consider sponsoring the project via GitHub Sponsors, Ko-fi or reach out directly to wing@openaccessaicollective.org.\n\n\n💎 Diamond Sponsors - Contact directly\n\n\n\n🥇 Gold Sponsors - $5000/mo\n\n\n\n🥈 Silver Sponsors - $1000/mo\n\n\n\n🥉 Bronze Sponsors - $500/mo\n\nJarvisLabs.ai", + "crumbs": [ + "Home" + ] + }, + { + "objectID": "examples/colab-notebooks/colab-axolotl-example.html", + "href": "examples/colab-notebooks/colab-axolotl-example.html", + "title": "Example notebook for running Axolotl on google colab", "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.", + "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)" + }, + { + "objectID": "examples/colab-notebooks/colab-axolotl-example.html#install-axolotl-and-dependencies", + "href": "examples/colab-notebooks/colab-axolotl-example.html#install-axolotl-and-dependencies", + "title": "Example notebook for running Axolotl on google colab", + "section": "Install Axolotl and dependencies", + "text": "Install Axolotl and dependencies\n\n!pip install torch==\"2.1.2\"\n!pip install -e git+https://github.com/axolotl-ai-cloud/axolotl#egg=axolotl\n!pip install flash-attn==\"2.5.0\"\n!pip install deepspeed==\"0.13.1\"!pip install mlflow==\"2.13.0\"" + }, + { + "objectID": "examples/colab-notebooks/colab-axolotl-example.html#create-an-yaml-config-file", + "href": "examples/colab-notebooks/colab-axolotl-example.html#create-an-yaml-config-file", + "title": "Example notebook for running Axolotl on google colab", + "section": "Create an yaml config file", + "text": "Create an yaml config file\n\nimport yaml\n\n# Your YAML string\nyaml_string = \"\"\"\nbase_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\nmodel_type: LlamaForCausalLM\ntokenizer_type: LlamaTokenizer\n\nload_in_8bit: false\nload_in_4bit: true\nstrict: false\n\ndatasets:\n - path: mhenrichsen/alpaca_2k_test\n type: alpaca\ndataset_prepared_path:\nval_set_size: 0.05\noutput_dir: ./outputs/qlora-out\n\nadapter: qlora\nlora_model_dir:\n\nsequence_len: 4096\nsample_packing: true\neval_sample_packing: false\npad_to_sequence_len: true\n\nlora_r: 32\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\nlora_target_linear: true\nlora_fan_in_fan_out:\n\nwandb_project:\nwandb_entity:\nwandb_watch:\nwandb_name:\nwandb_log_model:\n\ngradient_accumulation_steps: 4\nmicro_batch_size: 2\nnum_epochs: 4\noptimizer: paged_adamw_32bit\nlr_scheduler: cosine\nlearning_rate: 0.0002\n\ntrain_on_inputs: false\ngroup_by_length: false\nbf16: auto\nfp16:\ntf32: false\n\ngradient_checkpointing: true\nearly_stopping_patience:\nresume_from_checkpoint:\nlocal_rank:\nlogging_steps: 1\nxformers_attention:\nflash_attention: true\n\nwarmup_steps: 10\nevals_per_epoch: 4\nsaves_per_epoch: 1\ndebug:\ndeepspeed:\nweight_decay: 0.0\nfsdp:\nfsdp_config:\nspecial_tokens:\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)" + }, + { + "objectID": "examples/colab-notebooks/colab-axolotl-example.html#launch-the-training", + "href": "examples/colab-notebooks/colab-axolotl-example.html#launch-the-training", + "title": "Example notebook for running Axolotl on google colab", + "section": "Launch the training", + "text": "Launch the training\n\n# By using the ! the comand will be executed as a bash command\n!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml" + }, + { + "objectID": "examples/colab-notebooks/colab-axolotl-example.html#play-with-inference", + "href": "examples/colab-notebooks/colab-axolotl-example.html#play-with-inference", + "title": "Example notebook for running Axolotl on google colab", + "section": "Play with inference", + "text": "Play with inference\n\n# By using the ! the comand will be executed as a bash command\n!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n --qlora_model_dir=\"./qlora-out\" --gradio" + }, + { + "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": "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", - "NCCL" + "Multipack (Sample Packing)" ] }, { - "objectID": "docs/dataset_preprocessing.html", - "href": "docs/dataset_preprocessing.html", - "title": "Dataset Preprocessing", + "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": "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/multi-node.html", - "href": "docs/multi-node.html", - "title": "Multi Node", - "section": "", - "text": "You 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\nConfigure your model to use FSDP with for example:", + "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", - "Multi Node" + "Multipack (Sample Packing)" ] }, { - "objectID": "docs/multi-node.html#machine-configuration", - "href": "docs/multi-node.html#machine-configuration", - "title": "Multi Node", - "section": "Machine configuration", - "text": "Machine configuration\nOn each machine you need 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.\nOn the main machine only, make sure the port you set as main_process_port is open in TCP and reachable by other machines.\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.", + "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", - "Multi Node" + "Multipack (Sample Packing)" + ] + }, + { + "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/conversation.html", + "href": "docs/dataset-formats/conversation.html", + "title": "Conversation", + "section": "", + "text": "conversations where from is human/gpt. (optional: first row with role system to override default system prompt)\n\n\ndata.jsonl\n\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\n\nNote: type: sharegpt opens special configs: - conversation: enables conversions to many Conversation types. Refer to the ‘name’ here for options. - roles: allows you to specify the roles for input and output. This is useful for datasets with custom roles such as tool etc to support masking. - field_human: specify the key to use instead of human in the conversation. - field_model: specify the key to use instead of gpt in the conversation.\ndatasets:\n path: ...\n type: sharegpt\n\n conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n field_human: # Optional[str]. Human key to use for conversation.\n field_model: # Optional[str]. Assistant key to use for conversation.\n # Add additional keys from your dataset as input or output roles\n roles:\n input: # Optional[List[str]]. These will be masked based on train_on_input\n output: # Optional[List[str]].", + "crumbs": [ + "Dataset Formats", + "Conversation" + ] + }, + { + "objectID": "docs/dataset-formats/conversation.html#sharegpt", + "href": "docs/dataset-formats/conversation.html#sharegpt", + "title": "Conversation", + "section": "", + "text": "conversations where from is human/gpt. (optional: first row with role system to override default system prompt)\n\n\ndata.jsonl\n\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}\n\nNote: type: sharegpt opens special configs: - conversation: enables conversions to many Conversation types. Refer to the ‘name’ here for options. - roles: allows you to specify the roles for input and output. This is useful for datasets with custom roles such as tool etc to support masking. - field_human: specify the key to use instead of human in the conversation. - field_model: specify the key to use instead of gpt in the conversation.\ndatasets:\n path: ...\n type: sharegpt\n\n conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n field_human: # Optional[str]. Human key to use for conversation.\n field_model: # Optional[str]. Assistant key to use for conversation.\n # Add additional keys from your dataset as input or output roles\n roles:\n input: # Optional[List[str]]. These will be masked based on train_on_input\n output: # Optional[List[str]].", + "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#sharegpt.load_role", + "href": "docs/dataset-formats/conversation.html#sharegpt.load_role", + "title": "Conversation", + "section": "sharegpt.load_role", + "text": "sharegpt.load_role\nconversations where role is used instead of from\n\n\ndata.jsonl\n\n{\"conversations\": [{\"role\": \"...\", \"value\": \"...\"}]}", + "crumbs": [ + "Dataset Formats", + "Conversation" + ] + }, + { + "objectID": "docs/dataset-formats/conversation.html#sharegpt.load_guanaco", + "href": "docs/dataset-formats/conversation.html#sharegpt.load_guanaco", + "title": "Conversation", + "section": "sharegpt.load_guanaco", + "text": "sharegpt.load_guanaco\nconversations where from is prompter assistant instead of default sharegpt\n\n\ndata.jsonl\n\n{\"conversations\": [{\"from\": \"...\", \"value\": \"...\"}]}", + "crumbs": [ + "Dataset Formats", + "Conversation" + ] + }, + { + "objectID": "docs/dataset-formats/conversation.html#sharegpt_jokes", + "href": "docs/dataset-formats/conversation.html#sharegpt_jokes", + "title": "Conversation", + "section": "sharegpt_jokes", + "text": "sharegpt_jokes\ncreates a chat where bot is asked to tell a joke, then explain why the joke is funny\n\n\ndata.jsonl\n\n{\"conversations\": [{\"title\": \"...\", \"text\": \"...\", \"explanation\": \"...\"}]}", + "crumbs": [ + "Dataset Formats", + "Conversation" + ] + }, + { + "objectID": "docs/dataset-formats/index.html", + "href": "docs/dataset-formats/index.html", + "title": "Dataset Formats", + "section": "", + "text": "Axolotl supports a variety of dataset formats. It is recommended to use a JSONL format. The schema of the JSONL depends upon the task and the prompt template you wish to use. Instead of a JSONL, you can also use a HuggingFace dataset with columns for each JSONL field.\nBelow are these various formats organized by task:\n\n\n\n\n\n\n\n\n\nTitle\n\n\nDescription\n\n\n\n\n\n\nPre-training\n\n\nData format for a pre-training completion task.\n\n\n\n\nInstruction Tuning\n\n\nInstruction tuning formats for supervised fine-tuning.\n\n\n\n\nConversation\n\n\nConversation format for supervised fine-tuning.\n\n\n\n\nTemplate-Free\n\n\nConstruct prompts without a template.\n\n\n\n\nCustom Pre-Tokenized Dataset\n\n\nHow to use a custom pre-tokenized dataset.\n\n\n\n\n\nNo matching items", + "crumbs": [ + "Dataset Formats" + ] + }, + { + "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 unsloth from source and downgrade xformers as unsloth is incompatible with the most up to date libraries.\npip install --no-deps \"unsloth @ git+https://github.com/unslothai/unsloth.git\"\npip install --no-deps --force-reinstall xformers==0.0.26.post1\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/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 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, sharegpt, 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\n # Optional[str] fastchat conversation type, only used with type: sharegpt\n conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py\n field_human: # Optional[str]. Human key to use for conversation.\n field_model: # Optional[str]. Assistant key to use for conversation.\n # Add additional keys from your dataset as input or output roles\n roles:\n input: # Optional[List[str]]. These will be masked based on train_on_input\n output: # Optional[List[str]].\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# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\n# 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\n# Saves the desired chat template to the tokenizer_config.json for easier inferencing\n# Currently supports chatml and inst (mistral/mixtral)\nchat_template: chatml\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\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\nhf_mlflow_log_artifacts: # set to true to copy each saved checkpoint on each save to mlflow artifact registry\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\ntorch_compile: # 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, integers for every N steps. decimal for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\nsave_strategy: # Set to `\"no\"` to skip checkpoint saves\nsave_steps: # Leave empty to save at each epoch\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]\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# - 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/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/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/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.", + "crumbs": [ + "FAQ" ] }, { diff --git a/sitemap.xml b/sitemap.xml index 052fc3d51..db71ab762 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,94 +2,98 @@ https://axolotl-ai-cloud.github.io/axolotl/docs/debugging.html - 2024-07-18T18:54:51.444Z + 2024-07-19T04:47:20.888Z - https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/batch_vs_grad.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/mac.html - 2024-07-18T18:54:51.448Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/config.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/unsloth.html - 2024-07-18T18:54:51.448Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/index.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/template_free.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html - 2024-07-18T18:54:51.448Z - - - https://axolotl-ai-cloud.github.io/axolotl/FAQS.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/TODO.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/examples/colab-notebooks/colab-axolotl-example.html - 2024-07-18T18:54:51.448Z - - - https://axolotl-ai-cloud.github.io/axolotl/index.html - 2024-07-18T18:54:51.456Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/pretraining.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/tokenized.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/rlhf.html - 2024-07-18T18:54:51.448Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/fsdp_qlora.html - 2024-07-18T18:54:51.444Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/nccl.html - 2024-07-18T18:54:51.448Z - - - https://axolotl-ai-cloud.github.io/axolotl/docs/dataset_preprocessing.html - 2024-07-18T18:54:51.444Z + https://axolotl-ai-cloud.github.io/axolotl/docs/torchao.html + 2024-07-19T04:47:20.892Z https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html - 2024-07-18T18:54:51.448Z + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset_preprocessing.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/nccl.html + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/fsdp_qlora.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/rlhf.html + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/tokenized.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/pretraining.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/index.html + 2024-07-19T04:47:20.900Z + + + https://axolotl-ai-cloud.github.io/axolotl/examples/colab-notebooks/colab-axolotl-example.html + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/TODO.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/FAQS.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/template_free.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/index.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/unsloth.html + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/config.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/mac.html + 2024-07-19T04:47:20.892Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/batch_vs_grad.html + 2024-07-19T04:47:20.888Z + + + https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html + 2024-07-19T04:47:20.888Z https://axolotl-ai-cloud.github.io/axolotl/docs/input_output.html - 2024-07-18T18:54:51.448Z + 2024-07-19T04:47:20.892Z