integration-torchtune

[Zeyi-Lin]

Description

与torchtune微调框架的集成，已在Gemma-2b模型微调上试验成功。

使用方法：

1. 安装torchtune：

pip install torchtune

2. 下载gemma-2b预训练模型到本地：

from modelscope import snapshot_download
model_dir = snapshot_download('AI-ModelScope/gemma-2b', cache_dir="./",)

3. 创建2B_qlora_single_device.yaml：

# Config for multi-device QLoRA finetuning in lora_finetune_single_device.py
# using a gemma 2B model
#
# This config assumes that you've run the following command before launching
# this run:
#   tune download google/gemma-2b --hf-token <HF_TOKEN> --ignore-patterns ""
#
# To launch on a single device, run the following command from root:
#   tune run lora_finetune_single_device --config gemma/2B_qlora_single_device
#
# You can add specific overrides through the command line. For example
# to override the checkpointer directory while launching training
# you can run:
#   tune run lora_finetune_single_device --config gemma/2B_qlora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works only for training on single device.

# Tokenizer
tokenizer:
  _component_: torchtune.models.gemma.gemma_tokenizer
  path: ./AI-ModelScope/gemma-2b/tokenizer.model

# Dataset
dataset:
  _component_: torchtune.datasets.alpaca_dataset
seed: null
shuffle: True

# Model Arguments
model:
  _component_: torchtune.models.gemma.qlora_gemma_2b
  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
  apply_lora_to_mlp: True
  lora_rank: 64
  lora_alpha: 16

checkpointer:
  _component_: torchtune.utils.FullModelHFCheckpointer
  checkpoint_dir: ./AI-ModelScope/gemma-2b/
  checkpoint_files: [
    model-00001-of-00002.safetensors,
    model-00002-of-00002.safetensors,
  ]
  recipe_checkpoint: null
  output_dir: ./output/gemma-2b
  model_type: GEMMA
resume_from_checkpoint: False

optimizer:
  _component_: torch.optim.AdamW
  lr: 2e-5

lr_scheduler:
  _component_: torchtune.modules.get_cosine_schedule_with_warmup
  num_warmup_steps: 100

loss:
  _component_: torch.nn.CrossEntropyLoss

# Fine-tuning arguments
batch_size: 4
epochs: 3
max_steps_per_epoch: null
gradient_accumulation_steps: 4
compile: False

# Training env
device: cuda

# Memory management
enable_activation_checkpointing: True

# Reduced precision
dtype: bf16

# Logging
metric_logger:
  _component_: swanlab.integration.torchtune.SwanLabLogger
  project: "gemma-fintune"
  experiment_name: "gemma-2b"
  log_dir: ${output_dir}
output_dir: ./output/alpaca-gemma-lora
log_every_n_steps: 1
log_peak_memory_stats: False

# Show case the usage of pytorch profiler
# Set enabled to False as it's only needed for debugging training
profiler:
  _component_: torchtune.utils.setup_torch_profiler
  enabled: False

  #Output directory of trace artifacts
  output_dir: ${output_dir}/profiling_outputs

  #`torch.profiler.ProfilerActivity` types to trace
  cpu: True
  cuda: True

  #trace options passed to `torch.profiler.profile`
  profile_memory: False
  with_stack: False
  record_shapes: True
  with_flops: False

  # `torch.profiler.schedule` options:
  # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
  wait_steps: 5
  warmup_steps: 5
  active_steps: 2
  num_cycles: 1

4. 开始训练：

tune run lora_finetune_single_device --config 2B_qlora_single_device.yaml

Closes: #617