# 并行策略
parallel:
  fully_shard_parallel_size: auto
  fsdp_plan:
    apply_modules:
      - vision_tower
      - vision_tower.encoder.blocks.{*}
      - mm_projector
      - language_model.model
      - language_model.model.embed_tokens
      - language_model.model.layers.{*}
      - language_model.model.layers.{*}.mlp.experts
      - language_model.lm_head
    hook_modules:
      - language_model.model.layers.{*}
    num_to_forward_prefetch: 1
    num_to_backward_prefetch: 1
    param_dtype: bf16
    reduce_dtype: fp32
  ulysses_parallel_size: 1
  expert_parallel_size: 128
  ep_plan:
    apply_modules:
      - language_model.model.layers.{*}.mlp.experts

### 数据相关配置
data:
  dataset_param:
    dataset_type: huggingface
    #数据集属性
    attr:
      images: images
      messages: messages
      role_tag: role
      content_tag: content
      user_tag: user
      assistant_tag: assistant

    # 数据预处理
    preprocess_parameters:
      model_name_or_path: &HF_MODEL_LOAD_PATH mindspeed_mm/fsdp/models/kimik2_5
      trust_remote_code: true
      use_fast_tokenizer: true
      split_special_tokens: false
      image_max_pixels: 262144
      image_min_pixels: 1024
      video_max_pixels: 16384
      video_min_pixels: 0
      video_fps: 2.0
      video_maxlen: 64

    basic_parameters:
      cutoff_len: 1024
      template: kimi_k25
      enable_thinking: false
      train_on_prompt: false
      mask_history: false
      dataset_dir: /data
      dataset: &DATASET_PATH /data/mllm_format_llava_instruct_data.json
      cache_dir: ./cache_dir/
      overwrite_cache: false
      preprocessing_batch_size: 1000
      preprocessing_num_workers: 16
      max_samples: null

  # 数据加载
  dataloader_param:
    pin_memory: true
    shuffle: true
    dataloader_mode: sampler
    drop_last: true
    sampler_type: BaseRandomBatchSampler
    num_workers: 8
    collate_param:
      model_name: qwen3vl
      ignore_pad_token_for_loss: true
    enable_preload: false

# 模型配置
model:
  model_id: kimi_k25
  model_name_or_path: *HF_MODEL_LOAD_PATH
  trust_remote_code: true

# 优化特性配置
features:
  loss_cfg:
    loss_type: default
  enable_chunk_loss: true
  chunkloss_plan:
    apply_module: language_model.lm_head
    chunk_size: 1024
  recompute: true
  recompute_plan:
    apply_modules:
      - vision_tower.encoder.blocks.{*}
      - language_model.model.layers.{*}
  # activation offload 配置
  enable_activation_offload: false
  activation_offload_plan:
    apply_modules:
     - vision_tower.encoder.blocks.{*}
     - language_model.model.layers.{*}
  # chunkmbs配置
  enable_chunk_mbs: false
  chunkmbs_plan:
    apply_modules:
      - language_model.model.layers.{*}
    chunk_mbs: 1  # 这个表示的是chunk之后的micro batchsize
    batch_dim: 0
    chunk_arg_indexs: [ 0 ]
    chunk_kwarg_names: [ "position_embeddings", "position_ids", "rope_deltas", "attention_mask" ]

# 训练配置
training:
  micro_batch_size: 1
  gradient_accumulation_steps: 1
  seed: 42
  lr: 1.0e-5
  lr_decay_style: cosine
  lr_warmup_ratio: 0.1
  weight_decay: 0
  train_iters: 10000
  clip_grad: 0.0
  init_model_with_meta_device: true
  optimizer: adamw
  adam_fused: true
  save_interval: 10000
  use_deter_comp: false
  load_rank0_and_broadcast: false # 从rank0加载checkpoint并广播到其他rank
  plugin:
    - mindspeed_mm/fsdp/models/kimik2_5
    - mindspeed_mm/fsdp/data/datasets/huggingface

# 工具配置
tools:
  profile:
    enable: false
    profile_type: static
    ranks: [0]
    static_param:
      level: level1
      with_stack: false
      with_memory: false
      record_shapes: false
      with_cpu: true
      save_path: ./profiling
      start_step: 10
      end_step: 11
      data_simplification: false
      aic_metrics_type: PipeUtilization
  memory_profile:
      enable: false
      start_step: 1
      end_step: 2
      save_path: ./memory_snapshot
      dump_ranks: [0]
      stacks: all
      max_entries: null
      mem_info: false