quickstart instructions for starting from runpod (#5)

README.md

@@ -24,7 +24,97 @@ datasets:
 - Optionally Download some datasets, see [data/README.md](data/README.md)
 
 
-- Create a new or update the existing YAML config [config/pythia_1_2B_alpaca.yml](config/pythia_1_2B_alpaca.yml)
+- Create a new or update the existing YAML config [config/sample.yml](config/sample.yml)
+
+```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
+# 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
+# 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
+# whether you are training a 4-bit quantized model
+load_4bit: true
+# this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
+load_in_8bit: 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
+# 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
+# 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
+# lora hyperparameters
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+  - q_proj
+  - v_proj
+#  - k_proj
+#  - o_proj
+lora_fan_in_fan_out: false
+# wandb configuration if your're using it
+wandb_project:
+wandb_watch:
+wandb_run_id:
+wandb_log_model: checkpoint
+# where to save the finsihed model to
+output_dir: ./completed-model
+# training hyperparameters
+batch_size: 8
+micro_batch_size: 2
+num_epochs: 3
+warmup_steps: 100
+learning_rate: 0.00003
+# 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:
+# 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:
+```
+
 - Install python dependencies with ONE of the following:
 
     - `pip3 install -e .[int4]` (recommended)
@@ -54,3 +144,29 @@ use_cpu: false
 
 - 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`~~
+
+
+## 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
+source <(curl -s https://raw.githubusercontent.com/winglian/axolotl/main/scripts/setup-runpod.sh)
+```
+
+- Once the setup script completes
+```shell
+accelerate launch scripts/finetune.py configs/quickstart.yml
+```
+
+- 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
+```
+

configs/accelerate/default_config.yaml

@@ -0,0 +1,15 @@
+compute_environment: LOCAL_MACHINE
+distributed_type: 'NO'
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

configs/llama_7B_4bit.yml

@@ -4,7 +4,7 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: true
 datasets:
-  - path: vicgalle/alpaca-gpt4
+  - path: tatsu-lab/alpaca  # original alpaca dataset
     type: alpaca
 dataset_prepared_path: data/last_run_prepared
 val_set_size: 0.04
@@ -29,6 +29,7 @@ output_dir: ./lora-test
 batch_size: 8
 micro_batch_size: 2
 num_epochs: 3
+warmup_steps: 100
 learning_rate: 0.00003
 train_on_inputs: false
 group_by_length: false
@@ -37,5 +38,8 @@ tf32: true
 gradient_checkpointing: false
 early_stopping_patience: 3
 resume_from_checkpoint:
+auto_resume_from_checkpoints: true
 local_rank:
 load_4bit: true
+xformers_attention: true
+flash_attention:

configs/quickstart.yml

@@ -0,0 +1,45 @@
+base_model: decapoda-research/llama-7b-hf-int4
+base_model_config: decapoda-research/llama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+load_in_8bit: true
+datasets:
+  - path: tatsu-lab/alpaca  # original alpaca dataset
+    type: alpaca
+dataset_prepared_path: data/last_run_prepared
+val_set_size: 0.04
+adapter: lora
+lora_model_dir:
+sequence_len: 1024
+max_packed_sequence_len: 1024
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+  - q_proj
+  - v_proj
+#  - k_proj
+#  - o_proj
+lora_fan_in_fan_out: false
+wandb_project:
+wandb_watch:
+wandb_run_id:
+wandb_log_model: checkpoint
+output_dir: ./lora-test
+batch_size: 4
+micro_batch_size: 1
+num_epochs: 3
+warmup_steps: 100
+learning_rate: 0.00003
+train_on_inputs: false
+group_by_length: false
+bf16: true
+tf32: true
+gradient_checkpointing: false
+early_stopping_patience: 3
+resume_from_checkpoint:
+auto_resume_from_checkpoints: true
+local_rank:
+load_4bit: true
+xformers_attention: true
+flash_attention:

configs/sample.yml

@@ -0,0 +1,86 @@
+# 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
+# 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
+# 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
+# whether you are training a 4-bit quantized model
+load_4bit: true
+# this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
+load_in_8bit: 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
+# 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
+# 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
+# lora hyperparameters
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+  - q_proj
+  - v_proj
+#  - k_proj
+#  - o_proj
+lora_fan_in_fan_out: false
+# wandb configuration if your're using it
+wandb_project:
+wandb_watch:
+wandb_run_id:
+wandb_log_model: checkpoint
+# where to save the finsihed model to
+output_dir: ./completed-model
+# training hyperparameters
+batch_size: 8
+micro_batch_size: 2
+num_epochs: 3
+warmup_steps: 100
+learning_rate: 0.00003
+# 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:
+# 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:

requirements.txt

@@ -12,3 +12,5 @@ wandb
 flash-attn
 deepspeed
 einops
+xformers
+

scripts/finetune.py

@@ -225,7 +225,14 @@ def train(
         )
 
     logging.info("Starting trainer...")
-    trainer.train(resume_from_checkpoint=cfg.resume_from_checkpoint)
+    resume_from_checkpoint = cfg.resume_from_checkpoint
+    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+        possible_checkpoints = [str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")]
+        if len(possible_checkpoints) > 0:
+            sorted_paths = sorted(possible_checkpoints, key=lambda path: int(path.split('-')[-1]))
+            resume_from_checkpoint = sorted_paths[-1]
+            logging.info(f"Using Auto-resume functionality to start with checkpoint at {resume_from_checkpoint}")
+    trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
     if cfg.local_rank == 0:
         # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading

scripts/setup-runpod.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+
+export WANDB_MODE=offline
+export WANDB_CACHE_DIR=/workspace/data/wandb-cache
+mkdir -p $WANDB_CACHE_DIR
+
+mkdir -p /workspace/data/huggingface-cache/{hub,datasets}
+export HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
+export HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
+export TRANSFORMERS_CACHE="/workspace/data/huggingface-cache/hub"
+export NCCL_P2P_DISABLE=1
+
+nvidia-smi
+num_gpus=$(nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)
+gpu_indices=$(seq 0 $((num_gpus - 1)) | paste -sd "," -)
+export CUDA_VISIBLE_DEVICES=$gpu_indices
+echo "CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"
+
+apt-get update
+apt-get install -y build-essential ninja-build vim git-lfs
+git lfs install
+pip3 install --force-reinstall https://download.pytorch.org/whl/nightly/cu117/torch-2.0.0.dev20230301%2Bcu117-cp38-cp38-linux_x86_64.whl --index-url https://download.pytorch.org/whl/nightly/cu117
+if [ -z "${TORCH_CUDA_ARCH_LIST}" ]; then # only set this if not set yet
+    # this covers most common GPUs that the installed version of pytorch supports
+    # python -c "import torch; print(torch.cuda.get_arch_list())"
+    export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+fi
+
+cd /workspace/
+git clone https://github.com/winglian/axolotl.git
+cd axolotl
+pip install -e .[int4]
+mkdir -p ~/.cache/huggingface/accelerate/
+cp configs/accelerate/default_config.yml ~/.cache/huggingface/accelerate/default_config.yml

src/axolotl/utils/models.py

@@ -66,7 +66,10 @@ def load_model(
             from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
             from huggingface_hub import snapshot_download
 
-            cache_model_path = Path(snapshot_download(base_model))
+            snapshot_download_kwargs = {}
+            if cfg.base_model_ignore_patterns:
+                snapshot_download_kwargs["ignore_patterns"] = cfg.base_model_ignore_patterns
+            cache_model_path = Path(snapshot_download(base_model, ** snapshot_download_kwargs))
             files = (
                 list(cache_model_path.glob("*.pt"))
                 + list(cache_model_path.glob("*.safetensors"))

src/axolotl/utils/trainer.py

@@ -11,9 +11,9 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     total_num_steps = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
-    warmup_steps = min(int(0.03 * total_num_steps), 100)
+    warmup_steps = cfg.warmup_steps if cfg.warmup_steps else min(int(0.03 * total_num_steps), 100)
     logging_steps = max(min(int(0.005 * total_num_steps), 10), 1)
-    save_steps = eval_steps = min(int(0.05 * total_num_steps), 200)
+    save_steps = eval_steps = cfg.save_steps if cfg.save_steps else min(int(0.05 * total_num_steps), 200)
 
     training_arguments_kwargs = {}
     if cfg.bf16 == "full":
@@ -45,24 +45,23 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         **training_arguments_kwargs,
     )
 
-    decay_parameters = get_parameter_names(model, [nn.LayerNorm])
-    decay_parameters = [name for name in decay_parameters if "bias" not in name]
-    optimizer_grouped_parameters = [
-        {
-            "params": [p for n, p in model.named_parameters() if n in decay_parameters],
-            "weight_decay": training_args.weight_decay,
-        },
-        {
-            "params": [
-                p for n, p in model.named_parameters() if n not in decay_parameters
-            ],
-            "weight_decay": 0.0,
-        },
-    ]
-
     trainer_kwargs = {}
 
     if cfg.load_in_8bit and not cfg.load_4bit:
+        decay_parameters = get_parameter_names(model, [nn.LayerNorm])
+        decay_parameters = [name for name in decay_parameters if "bias" not in name]
+        optimizer_grouped_parameters = [
+            {
+                "params": [p for n, p in model.named_parameters() if n in decay_parameters],
+                "weight_decay": training_args.weight_decay,
+            },
+            {
+                "params": [
+                    p for n, p in model.named_parameters() if n not in decay_parameters
+                ],
+                "weight_decay": 0.0,
+            },
+        ]
         optimizer = bnb.optim.Adam8bit(
             optimizer_grouped_parameters,
             betas=(training_args.adam_beta1, training_args.adam_beta2),