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

[fix] fix determine_n_share_experts_fusion (#6118)

Browse Source
This commit is contained in:
JieXin Liang
2025-05-10 16:19:09 +08:00
committed by GitHub
parent b29a026e14
commit c178abdabc
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/models/deepseek_v2.py
View File

@@ -1486,14 +1486,15 @@ class DeepseekV2ForCausalLM(nn.Module):
if self.n_share_experts_fusion > 0:
# Only Deepseek V3/R1 can use shared experts fusion optimization now.
if (
self.config.architectures[0] != architecture
not _is_cuda
or self.config.architectures[0] != architecture
or self.config.n_routed_experts != 256
):
self.n_share_experts_fusion = 0
global_server_args_dict["n_share_experts_fusion"] = 0
log_info_on_rank0(
logger,
"Only Deepseek V3/R1 can use shared experts fusion optimization. Shared experts fusion optimization is disabled.",
"Only Deepseek V3/R1 on NV-platform can use shared experts fusion optimization. Shared experts fusion optimization is disabled.",
)
else:
assert (
@@ -1501,7 +1502,8 @@ class DeepseekV2ForCausalLM(nn.Module):
), f"Shared experts fusion optimization is enabled in DeepSeek V3/R1, set it to {self.tp_size} can get best optimized performace."
elif self.n_share_experts_fusion == 0:
if (
torch.cuda.get_device_capability("cuda") >= (9, 0)
_is_cuda
and torch.cuda.get_device_capability("cuda") >= (9, 0)
and self.config.architectures[0] == architecture
and self.config.n_routed_experts == 256
and (not global_server_args_dict["enable_deepep_moe"])