model_config:
  model_name: longcat_flash
  model_path: /data/models/origin/LongCat-Flash-Chat-W8A8/
  with_ckpt: True                   # [False, True]
  exe_mode: "eager"
  enable_static_kernel: False         # [False, True] Only takes effect in npugraph_ex mode.
  enable_profiler: False            # [False, True]
  force_eplb: False                 # [False, True]
  enable_cache_compile: False       # [False, True]
  custom_params:
    enable_mla_prolog: True           # [False, True]
    enable_multi_streams: 0           # [0, 1, 2]
    enable_prefetch: False            # [False, True]
    enable_superkernel: False         # [False, True]
    moe_chunk_max_len: 1024

parallel_config:
  world_size: 16
  attn_tp_size: 16
  dense_tp_size: 16
  moe_tp_size: 1
  embed_tp_size: 16
  lmhead_tp_size: 16
  o_proj_tp_size: 16

scheduler_config:
  block_size: 128
  max_new_tokens: 1
  max_prefill_tokens: 8192
  batch_size: 16
  mem_fraction_static: 0.85
  num_reserved_decode_tokens: 512