HF_MODEL_LOAD_PATH: &HF_MODEL_LOAD_PATH ./ckpt/hf_path/Qwen3-VL-30B-A3B-Instruct
MM_MODEL_LOAD_PATH: &MM_MODEL_LOAD_PATH ./ckpt/mm_path/Qwen3-VL-30B-A3B-Instruct
DATASET_PATH: &DATASET_PATH ./data/mllm_format_llava_instruct_data.json
FSDP2_PATH: &FSDP2_PATH ./examples/qwen3vl/fsdp2_config.yaml
SAVE_PATH: &SAVE_PATH save_dir
MM_TOOL_PATH: ./mindspeed_mm/tools/tools.json


### GPT args
gpt_args:
  ## parallel: （TP/PP/CP/FSDP）
  context_parallel_size: 1
  context_parallel_algo: ulysses_cp_algo
  use_torch_fsdp2: true
  fsdp2_config_path: *FSDP2_PATH
  ckpt_format: torch_dcp

  ## training:
  clip_grad: 0.0
  micro_batch_size: 1
  global_batch_size: 8  # global_batch_sizes=micro_batch_size*grad_acc_step*dp=micro_batch_size*grad_acc_step*(world_size/cp)
  optimizer_selection: fused_torch_adamw
  lr: 1.0e-5
  weight_decay: 0.0
  adam_beta1: 0.9
  adam_beta2: 0.999
  lr_decay_style: cosine
  lr_warmup_fraction: 0.1
  train_iters: 10000
  no_load_optim: true  # 不加载优化器状态，若需加载请移除
  no_load_rng: true  # 不加载随机数状态，若需加载请移除
  no_save_optim: true  # 不保存优化器状态，若需保存请移除
  no_save_rng: true  # 不保存随机数状态，若需保存请移除
  seed: 42

  ## save_and_logging:
  log_interval: 1
  save_interval: 10000
  eval_interval: 10000
  eval_iters: 5000
  save: *SAVE_PATH
  log_tps: true

  ## model:
  load: *MM_MODEL_LOAD_PATH
  use_cpu_initialization: true
  init_model_with_meta_device: true
  normalization: RMSNorm
  use_fused_rmsnorm: true
  swiglu: true
  use_fused_swiglu: true
  use_flash_attn: true
  vocab_size: 152064
  seq_length: 1024
  make_vocab_size_divisible_by: 1
  tokenizer_type: NullTokenizer
  no_gradient_accumulation_fusion: true
  untie_embeddings_and_output_weights: true
  no_masked_softmax_fusion: true

  ## data:
  num_workers: 8


# LoRA args
lora_args:
  lora_r: 8
  lora_alpha: 16
  lora_dropout: 0.0
  lora_target_modules: v_proj,o_proj,k_proj,q_proj
  lora_register_forward_hook: embed_tokens


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

    # 数据预处理
    preprocess_parameters:
      model_name_or_path: *HF_MODEL_LOAD_PATH
      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:
      template: qwen3_vl_nothink
      enable_thinking: false
      train_on_prompt: false
      mask_history: false
      tool_format: null
      dataset_dir: ./data
      dataset: *DATASET_PATH
      cache_dir: ./data/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
    collate_param:
      model_name: qwen3vl
      ignore_pad_token_for_loss: true


### 模型相关配置
model:
  model_id: qwen3_vl_moe
  init_from_hf_path: *HF_MODEL_LOAD_PATH
  image_encoder:
    vision_encoder:
      model_id: qwen3vit
      num_layers: 27  # 此参数为megatron校验参数，运行中不生效，如需修改请修改HF_MODEL_LOAD_PATH目录下的config.json
      hidden_size: 1152  # 此参数为megatron校验参数，运行中不生效
      num_attention_heads: 16  # 此参数为megatron校验参数，运行中不生效
      freeze: true
      attn_implementation: flash_attention_2
      attn_layout: TND
      synchronize_per_layer: true
      recompute_granularity: full
    vision_projector:
      model_id: lnmlp
      num_layers: 1
      freeze: true
  text_decoder:
    model_id: qwen3lm
    num_layers: 48  # 此参数为megatron校验参数，运行中不生效，如需修改请修改HF_MODEL_LOAD_PATH目录下的config.json
    hidden_size: 2048  # 此参数为megatron校验参数，运行中不生效
    num_attention_heads: 32  # 此参数为megatron校验参数，运行中不生效
    max_position_embeddings: 262144  # 此参数为megatron校验参数，运行中不生效
    freeze: false
    use_npu_fused_moe: false
    attn_implementation: flash_attention_2
    attn_layout: TND
    is_causal: false
    activation_offload: false
    synchronize_per_layer: true
    recompute_granularity: full
  loss_cfg:
      compute_mode: default
      chunk_size: 1024
      router_aux_loss_coef: 0.0
      loss_type: default
  patch:
      clip_grad_async: true
      bridge_patch: false