veomni_cli.yaml

data:
  chat_template: default
  data_tag: default
  data_type: plaintext
  dataloader_type: native
  datasets_type: iterable
  drop_last: true
  drop_resume_buffer: false
  enable_multisource: false
  image_keys: images
  max_seq_len: 1024
  multisource_datasets_type: interleave
  num_workers: 2
  pin_memory: true
  predownload_factor: 0.5
  prefetch_factor: 2
  shuffle_shard_nums: 1
  silent_exception: false
  source_name: null
  split_nums: 1
  text_keys: text
  train_path: /scratch/labeling/redstone/warc
  train_size: 1000000000000
model:
  attn_implementation: flash_attention_2
  basic_modules: []
  config_path: /scratch/data/30bmerge
  decoders: {}
  encode_target: false
  encoders: {}
  foundation: {}
  input_encoder: encoder
  model_path: /scratch/data/30bmerge
  moe_implementation: fused
  output_encoder: decoder
  tokenizer_path: /scratch/data/30bmerge
train:
  activation_gpu_limit: 0.0
  async_enabled: false
  broadcast_model_weights_from_rank0: true
  bsz_warmup_init_mbtoken: 200
  bsz_warmup_ratio: 0.007
  ckpt_manager: dcp
  context_parallel_size: 1
  data_parallel_mode: fsdp1
  data_parallel_replicate_size: 1
  data_parallel_shard_size: 8
  dyn_bsz: true
  dyn_bsz_buffer_size: 200
  dyn_bsz_margin: 0
  dyn_bsz_runtime: worker
  empty_cache_steps: 500
  enable_activation_offload: false
  enable_compile: false
  enable_forward_prefetch: true
  enable_fsdp_offload: false
  enable_full_determinism: false
  enable_full_shard: true
  enable_gradient_checkpointing: true
  enable_mixed_precision: false
  enable_profiling: false
  enable_reentrant: false
  ep_outside: false
  expert_parallel_size: 8
  gc_steps: 500
  global_batch_size: 32
  init_device: cuda
  load_checkpoint_path: ''
  lr: 0.0003
  lr_decay_ratio: 1.0
  lr_decay_style: constant
  lr_min: 1.0e-07
  lr_start: 0.0
  lr_warmup_ratio: 0.007
  max_grad_norm: 1.0
  max_steps: null
  micro_batch_size: 1
  no_decay_modules: []
  no_decay_params: []
  num_train_epochs: 1
  optimizer: adamw
  output_dir: output_q330b_v3
  pipeline_parallel_size: 1
  profile_end_step: 2
  profile_profile_memory: true
  profile_rank0_only: true
  profile_record_shapes: true
  profile_start_step: 1
  profile_trace_dir: ./trace
  profile_with_stack: true
  rmpad: false
  rmpad_with_pos_ids: true
  save_async: false
  save_epochs: 1
  save_hf_weights: true
  save_steps: 100
  seed: 42
  tensor_parallel_size: 1
  train_architecture: full
  ulysses_parallel_size: 1
  use_wandb: false
  vit_lr: 5.0e-05
  wandb_name: Qwen3-30B-A3B-CT
  wandb_project: Qwen3-30B-A3B
  weight_decay: 0.01