TensorRT-LLM/examples/auto_deploy/model_registry/configs/qwen3.5_moe_400b.yaml at main · ixlmar/TensorRT-LLM · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
runtime: trtllm
compile_backend: torch-cudagraph
max_seq_len: 2048
max_num_tokens: 2048
max_batch_size: 512
cuda_graph_batch_sizes: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512]
world_size: 8
enable_chunked_prefill: true
model_factory: AutoModelForCausalLM
kv_cache_config:
  enable_block_reuse: false
  free_gpu_memory_fraction: 0.95
  tokens_per_block: 64
model_kwargs:
  torch_dtype: bfloat16
transforms:
  export_to_gm:
    num_moe_experts_for_export: 2
  fuse_gemms_mixed_children:
    enabled: true
  detect_sharding:
    sharding_dims: ['tp','ep', 'bmm']
    # use only manual config for TP sharding
    sharding_source: ['manual']
    manual_config:
      tp_plan:
        # GDN layer
        "in_proj_qkv": "delta"
        # attention layer
        "q_proj": "colwise"
        "k_proj": "colwise"
        "v_proj": "colwise"
        "o_proj": "rowwise"
        # replicating shared experts (keep them commented out)
        # "shared_expert_gate_proj": "colwise"
        # "shared_expert_up_proj": "colwise"
        # "shared_expert_down_proj": "rowwise"
        # gating layer should be replicated as well
        # "gate": "gather"