yandex/YandexGPT-5-Lite-8B-pretrain · Дообучение модели через torchtune

Вот пример базового рецепта по дообучению с lora:

Скачиваем репозиторий:

mkdir torchtune
cd torchtune
tune download yandex/YandexGPT-5-Lite-8B-pretrain \
  --output-dir YandexGPT-5-Lite-8B-pretrain

Смотрим список конфигов и копируем подходящий под задачу:

tune ls
tune cp llama3_1/8B_lora training_config.yaml

Изменяем конфиг – адаптируем его под нашу модель и делаем подходящим под задачу. Например, такой вариант подойдет для lora-обучения на инстракт датасете alpaca-cleaned:

output_dir: <HOME_PATH>/torchtune/result_dir # /home/vorobyov01/torchtune/result_dir

# Tokenizer
tokenizer:
  _component_: torchtune.models.llama2.llama2_tokenizer
  path: <HOME_PATH>/torchtune/YandexGPT-5-Lite-8B-pretrain/tokenizer.model
  max_seq_len: null

# Model Arguments
model:
  _component_: torchtune.models.llama3.lora_llama3
  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
  apply_lora_to_mlp: True
  apply_lora_to_output: False
  lora_rank: 8
  lora_alpha: 16
  lora_dropout: 0.0

  vocab_size: 129024
  num_layers: 32
  num_heads: 32
  num_kv_heads: 8
  embed_dim: 4096
  max_seq_len: 32768
  intermediate_dim: 14336
  attn_dropout: 0.0
  norm_eps: 0.000001

checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: <HOME_PATH>/torchtune/YandexGPT-5-Lite-8B-pretrain
  checkpoint_files: [
    model-00001-of-00004.safetensors,
    model-00002-of-00004.safetensors,
    model-00003-of-00004.safetensors,
    model-00004-of-00004.safetensors
  ]
  recipe_checkpoint: null
  output_dir: ${output_dir}
  model_type: LLAMA3
resume_from_checkpoint: False
save_adapter_weights_only: False

# Dataset and Sampler
dataset:
  _component_: torchtune.datasets.alpaca_cleaned_dataset
  packed: False  # True increases speed
seed: null
shuffle: True
batch_size: 2

# Optimizer and Scheduler
optimizer:
  _component_: torch.optim.AdamW
  fused: True
  weight_decay: 0.01
  lr: 3e-4
lr_scheduler:
  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
  num_warmup_steps: 100

loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss

# Training
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 8  # Use to increase effective batch size
compile: False  # torch.compile the model + loss, True increases speed + decreases memory

# Logging
metric_logger:
  _component_: torchtune.training.metric_logging.DiskLogger
  log_dir: ${output_dir}/logs
log_every_n_steps: 1
log_peak_memory_stats: True

# Environment
device: cuda
dtype: bf16
enable_activation_checkpointing: False  # True reduces memory
enable_activation_offloading: False  # True reduces memory


# Profiler (disabled)
profiler:
  _component_: torchtune.training.setup_torch_profiler
  enabled: False
12
  #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: 3
  active_steps: 2
  num_cycles: 1

Запускаем обучение:

tune run lora_finetune_single_device --config training_config.yaml

Можно на нескольких GPU:

CUDA_VISIBLE_DEVICES="0,1,2,3" tune run --nproc-per-node 4 lora_finetune_distributed --config training_config.yaml