chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

support mtp with deepseek r1 nvfp4 model (#13115)

Browse Source 
Co-authored-by: Trevor Morris <tmorris@nvidia.com>
This commit is contained in:
Rain Jiang
2025-12-06 00:45:54 -08:00
committed by GitHub
parent 42fcf5438f
commit ea177372bd
11 changed files with 123 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/server_args.py
View File

@@ -168,6 +168,8 @@ MOE_RUNNER_BACKEND_CHOICES = [
"cutlass",
]
MOE_A2A_BACKEND_CHOICES = ["none", "deepep", "mooncake", "ascend_fuseep"]
MAMBA_SSM_DTYPE_CHOICES = ["float32", "bfloat16"]
@@ -395,6 +397,7 @@ class ServerArgs:
speculative_token_map: Optional[str] = None
speculative_attention_mode: str = "prefill"
speculative_moe_runner_backend: Optional[str] = None
speculative_moe_a2a_backend: Optional[str] = None
# Speculative decoding (ngram)
speculative_ngram_min_match_window_size: int = 1
@@ -3011,6 +3014,13 @@ class ServerArgs:
default=ServerArgs.speculative_moe_runner_backend,
help="Choose the runner backend for MoE in speculative decoding.",
)
parser.add_argument(
"--speculative-moe-a2a-backend",
type=str,
choices=MOE_A2A_BACKEND_CHOICES,
default=ServerArgs.speculative_moe_a2a_backend,
help="Choose the backend for MoE A2A in speculative decoding",
)
# Speculative decoding (ngram)
parser.add_argument(
@@ -3069,7 +3079,7 @@ class ServerArgs:
parser.add_argument(
"--moe-a2a-backend",
type=str,
choices=["none", "deepep", "mooncake", "ascend_fuseep"],
choices=MOE_A2A_BACKEND_CHOICES,
default=ServerArgs.moe_a2a_backend,
help="Choose the backend for MoE A2A.",
)