EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Refactor] Rename n_share_experts_fusion as num_fused_shared_experts (#6735)

Browse Source
This commit is contained in:
Cheng Wan
2025-06-03 17:48:24 -07:00
committed by GitHub
parent b6d0ce9f78
commit 8a5480528d
14 changed files with 82 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/server_args.py
View File

@@ -206,7 +206,7 @@ class ServerArgs:
flashinfer_mla_disable_ragged: bool = False
warmups: Optional[str] = None
moe_dense_tp_size: Optional[int] = None
n_share_experts_fusion: int = 0
num_fused_shared_experts: int = 0
disable_chunked_prefix_cache: bool = False
disable_fast_image_processor: bool = False
mm_attention_backend: Optional[str] = None
@@ -1373,11 +1373,11 @@ class ServerArgs:
)
parser.add_argument(
"--n-share-experts-fusion",
"--num-fused-shared-experts",
type=int,
default=0,
help="The number of shared_experts need to be replicated to fuse with normal experts in deepseek v3/r1, "
"set it to tp_size can get best optimized performance. Note that for architectures with SM==90, we have enabled the shared experts fusion optimization by default for DeepSeek V3/R1, with n_share_experts_fusion automatically set to the TP size.",
"set it to tp_size can get best optimized performance. Note that for architectures with SM==90, we have enabled the shared experts fusion optimization by default for DeepSeek V3/R1, with num_fused_shared_experts automatically set to the TP size.",
)
parser.add_argument(
"--disable-chunked-prefix-cache",