diff --git a/README.md b/README.md index 55689696e..1e46944db 100644 --- a/README.md +++ b/README.md @@ -1,71 +1,211 @@ # Axolotl -#### Go ahead and axolotl questions +
+ axolotl +
+

+ One repo to finetune them all! +

+

+ Go ahead and axolotl questions!! +

+
+
-## Support Matrix +## Axolotl supports | | fp16/fp32 | fp16/fp32 w/ lora | 4bit-quant | 4bit-quant w/flash attention | flash attention | xformers attention | |----------|:----------|:------------------|------------|------------------------------|-----------------|--------------------| | llama | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | | Pythia | ✅ | ✅ | ❌ | ❌ | ❌ | ❓ | | cerebras | ✅ | ✅ | ❌ | ❌ | ❌ | ❓ | +| mpt | ✅ | ❌ | ❌ | ❌ | ❌ | ❓ | -## Getting Started -- install python 3.9. 3.10 and above are not supported. +## Quickstart ⚡ -- Point the config you are using to a huggingface hub dataset (see [configs/llama_7B_4bit.yml](https://github.com/winglian/axolotl/blob/main/configs/llama_7B_4bit.yml#L6-L8)) +**Requirements**: Python 3.9. -```yaml -datasets: - - path: vicgalle/alpaca-gpt4 - type: alpaca +```bash +git clone https://github.com/OpenAccess-AI-Collective/axolotl + +pip3 install -e .[int4] + +accelerate config + +# finetune +accelerate launch scripts/finetune.py examples/4bit-lora-7b/config.yml + +# inference +accelerate launch scripts/finetune.py examples/4bit-lora-7b/config.yml \ + --inference --lora_model_dir="./llama-7b-lora-int4" ``` -- Optionally Download some datasets, see [data/README.md](data/README.md) +## Installation +### Environment -- Create a new or update the existing YAML config [config/sample.yml](config/sample.yml) +- Docker + ```bash + docker run --gpus '"all"' --rm -it winglian/axolotl:main + ``` + - `winglian/axolotl:dev`: dev branch + - `winglian/axolotl-runpod:main`: for runpod + +- Conda/Pip venv + 1. Install python **3.9** + + 2. Install python dependencies with ONE of the following: + - `pip3 install -e .[int4]` (recommended) + - `pip3 install -e .[int4_triton]` + - `pip3 install -e .` + +### Dataset + +Have dataset(s) in one of the following format (JSONL recommended): + +- `alpaca`: instruction; input(optional) + ```json + {"instruction": "...", "input": "...", "output": "..."} + ``` +- `sharegpt`: conversations + ```json + {"conversations": [{"from": "...", "value": "..."}]} + ``` +- `completion`: raw corpus + ```json + {"text": "..."} + ``` + +
+ +See other formats + +- `jeopardy`: question and answer + ```json + {"question": "...", "category": "...", "answer": "..."} + ``` +- `oasst`: instruction + ```json + {"INSTRUCTION": "...", "RESPONSE": "..."} + ``` +- `gpteacher`: instruction; input(optional) + ```json + {"instruction": "...", "input": "...", "response": "..."} + ``` +- `reflection`: instruction with reflect; input(optional) + ```json + {"instruction": "...", "input": "...", "output": "...", "reflection": "...", "corrected": "..."} + ``` + +> Have some new format to propose? Check if it's already defined in [data.py](src/axolotl/utils/data.py) in `dev` branch! + +
+ +Optionally, download some datasets, see [data/README.md](data/README.md) + +### Config + +See sample configs in [configs](configs) folder or [examples](examples) for quick start. It is recommended to duplicate and modify to your needs. The most important options are: + +- model + ```yaml + base_model: ./llama-7b-hf # local or huggingface repo + ``` + Note: The code will load the right architecture. + +- dataset + ```yaml + datasets: + - path: vicgalle/alpaca-gpt4 # local or huggingface repo + type: alpaca # format from earlier + sequence_len: 2048 # max token length / prompt + ``` + +- loading + ```yaml + load_4bit: true + load_in_8bit: true + bf16: true + fp16: true + tf32: true + ``` + Note: Repo does not do 4-bit quantization. + +- lora + ```yaml + adapter: lora # blank for full finetune + lora_r: 8 + lora_alpha: 16 + lora_dropout: 0.05 + lora_target_modules: + - q_proj + - v_proj + ``` + +
+ +All yaml options ```yaml # this is the huggingface model that contains *.pt, *.safetensors, or *.bin files # this can also be a relative path to a model on disk -base_model: decapoda-research/llama-7b-hf-int4 +base_model: ./llama-7b-hf # you can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc) base_model_ignore_patterns: # if the base_model repo on hf hub doesn't include configuration .json files, # you can set that here, or leave this empty to default to base_model -base_model_config: decapoda-research/llama-7b-hf +base_model_config: ./llama-7b-hf # If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too model_type: AutoModelForCausalLM # Corresponding tokenizer for the model AutoTokenizer is a good choice tokenizer_type: AutoTokenizer +# Trust remote code for untrusted source +trust_remote_code: + # whether you are training a 4-bit quantized model load_4bit: true +gptq_groupsize: 128 # group size +gptq_model_v1: false # v1 or v2 + # this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer load_in_8bit: true + +# Use CUDA bf16 +bf16: true +# Use CUDA fp16 +fp16: true +# Use CUDA tf32 +tf32: true + # a list of one or more datasets to finetune the model with datasets: # this can be either a hf dataset, or relative path - path: vicgalle/alpaca-gpt4 # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection] type: alpaca + data_files: # path to source data files + # axolotl attempts to save the dataset as an arrow after packing the data together so # subsequent training attempts load faster, relative path dataset_prepared_path: data/last_run_prepared +# push prepared dataset to hub +push_dataset_to_hub: # repo path # How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc val_set_size: 0.04 -# if you want to use lora, leave blank to train all parameters in original model -adapter: lora -# if you already have a lora model trained that you want to load, put that here -lora_model_dir: + # the maximum length of an input to train with, this should typically be less than 2048 # as most models have a token/context limit of 2048 sequence_len: 2048 # max sequence length to concatenate training samples together up to # inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning max_packed_sequence_len: 1024 + +# if you want to use lora, leave blank to train all parameters in original model +adapter: lora +# if you already have a lora model trained that you want to load, put that here # lora hyperparameters +lora_model_dir: lora_r: 8 lora_alpha: 16 lora_dropout: 0.05 @@ -74,14 +214,24 @@ lora_target_modules: - v_proj # - k_proj # - o_proj +# - gate_proj +# - down_proj +# - up_proj +lora_modules_to_save: +# - embed_tokens +# - lm_head +lora_out_dir: lora_fan_in_fan_out: false -# wandb configuration if your're using it + +# wandb configuration if you're using it wandb_project: wandb_watch: wandb_run_id: -wandb_log_model: checkpoint -# where to save the finsihed model to +wandb_log_model: # 'checkpoint' + +# where to save the finished model to output_dir: ./completed-model + # training hyperparameters batch_size: 8 micro_batch_size: 2 @@ -89,87 +239,110 @@ eval_batch_size: 2 num_epochs: 3 warmup_steps: 100 learning_rate: 0.00003 +logging_steps: + # whether to mask out or include the human's prompt from the training labels train_on_inputs: false # don't use this, leads to wonky training (according to someone on the internet) group_by_length: false -# Use CUDA bf16 -bf16: true -# Use CUDA tf32 -tf32: true + # does not work with current implementation of 4-bit LoRA gradient_checkpointing: false + # stop training after this many evaluation losses have increased in a row # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback early_stopping_patience: 3 # specify a scheduler to use with the optimizer. only one_cycle is supported currently lr_scheduler: +# specify optimizer +optimizer: +# specify weight decay +weight_decay: + # whether to use xformers attention patch https://github.com/facebookresearch/xformers: xformers_attention: # whether to use flash attention patch https://github.com/HazyResearch/flash-attention: flash_attention: + # resume from a specific checkpoint dir resume_from_checkpoint: # if resume_from_checkpoint isn't set and you simply want it to start where it left off # be careful with this being turned on between different models auto_resume_from_checkpoints: false + # don't mess with this, it's here for accelerate and torchrun local_rank: + +# add or change special tokens +special_tokens: + # bos_token: "" + # eos_token: "" + # unk_token: "" +# add extra tokens +tokens: + +# FSDP +fsdp: +fsdp_config: + +# Deepspeed +deepspeed: + +# TODO +torchdistx_path: + +# Debug mode +debug: ``` -- Install python dependencies with ONE of the following: +
- - `pip3 install -e .[int4]` (recommended) - - `pip3 install -e .[int4_triton]` - - `pip3 install -e .` -- -- If not using `int4` or `int4_triton`, run `pip install "peft @ git+https://github.com/huggingface/peft.git"` -- Configure accelerate `accelerate config` or update `~/.cache/huggingface/accelerate/default_config.yaml` +### Accelerate -```yaml -compute_environment: LOCAL_MACHINE -distributed_type: MULTI_GPU -downcast_bf16: 'no' -gpu_ids: all -machine_rank: 0 -main_training_function: main -mixed_precision: bf16 -num_machines: 1 -num_processes: 4 -rdzv_backend: static -same_network: true -tpu_env: [] -tpu_use_cluster: false -tpu_use_sudo: false -use_cpu: false +Configure accelerate + +```bash +accelerate config + +# Edit manually +# nano ~/.cache/huggingface/accelerate/default_config.yaml ``` -- Train! `accelerate launch scripts/finetune.py`, make sure to choose the correct YAML config file -- Alternatively you can pass in the config file like: `accelerate launch scripts/finetune.py configs/llama_7B_alpaca.yml`~~ +### Train - -## How to start training on Runpod in under 10 minutes - -- Choose your Docker container wisely. -- I recommend `huggingface:transformers-pytorch-deepspeed-latest-gpu` see https://hub.docker.com/r/huggingface/transformers-pytorch-deepspeed-latest-gpu/ -- Once you start your runpod, and SSH into it: -```shell -export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX" -source <(curl -s https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/dev/scripts/setup-runpod.sh) +Run +```bash +accelerate launch scripts/finetune.py configs/your_config.yml ``` -- Once the setup script completes -```shell -accelerate launch scripts/finetune.py configs/quickstart.yml +### Inference + +Add `--inference` flag to train command above + +If you are inferencing a pretrained LORA, pass +```bash +--lora_model_dir ./completed-model ``` -- Here are some helpful environment variables you'll want to manually set if you open a new shell -```shell -export WANDB_MODE=offline -export WANDB_CACHE_DIR=/workspace/data/wandb-cache -export HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets" -export HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub" -export TRANSFORMERS_CACHE="/workspace/data/huggingface-cache/hub" -export NCCL_P2P_DISABLE=1 +### Merge LORA to base (Dev branch 🔧 ) + +Add below flag to train command above + +```bash +--merge_lora --lora_model_dir="./completed-model" ``` +## Common Errors 🧰 + +> Cuda out of memory + +Please reduce any below + - `micro_batch_size` + - `eval_batch_size` + - `sequence_len` + +## Contributing 🤝 + +Bugs? Please check for open issue else create a new [Issue](https://github.com/OpenAccess-AI-Collective/axolotl/issues/new). + +PRs are **greatly welcome**! \ No newline at end of file diff --git a/data/README.md b/data/README.md index 57b8b76c8..34d7a5659 100644 --- a/data/README.md +++ b/data/README.md @@ -1,6 +1,5 @@ -- Download some datasets -- +## Download some datasets ```shell curl https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_gpt4.json -o data/raw/alpaca_data_gpt4.json curl https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json -L -o data/raw/vicuna_cleaned.json @@ -8,7 +7,7 @@ curl https://github.com/teknium1/GPTeacher/blob/main/Instruct/gpt4-instruct-simi curl https://github.com/teknium1/GPTeacher/blob/main/Roleplay/roleplay-similarity_0.6-instruct-dataset.json?raw=true -L -o data/raw/roleplay-similarity_0.6-instruct-dataset.json ``` -- Convert the JSON data files to JSONL. +## Convert the JSON data files to JSONL. ```shell python3 ./scripts/alpaca_json_to_jsonl.py --input data/alpaca_data_gpt4.json > data/alpaca_data_gpt4.jsonl @@ -16,8 +15,9 @@ python3 ./scripts/alpaca_json_to_jsonl.py --input data/raw/vicuna_cleaned.json > python3 ./scripts/alpaca_json_to_jsonl.py --input data/raw/roleplay-similarity_0.6-instruct-dataset.json > data/roleplay-similarity_0.6-instruct-dataset.jsonl python3 ./scripts/alpaca_json_to_jsonl.py --input data/raw/gpt4-instruct-similarity-0.6-dataset.json > data/gpt4-instruct-similarity-0.6-dataset.jsonl ``` +--- -- Using JSONL makes it easier to subset the data if you want a smaller training set, i.e get 2000 random examples. +Using JSONL makes it easier to subset the data if you want a smaller training set, i.e get 2000 random examples. ```shell shuf -n2000 data/vicuna_cleaned.jsonl > data/vicuna_cleaned.subset0.jsonl diff --git a/image/axolotl.png b/image/axolotl.png new file mode 100644 index 000000000..21c27db85 Binary files /dev/null and b/image/axolotl.png differ