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

Fix judgment condition for enabling Deepseek V3/R1 shared expert fusion optimization (#7371)

Browse Source
This commit is contained in:
Li Hui
2025-06-20 12:58:00 +08:00
committed by GitHub
parent 97011abc8a
commit a06912ad8b
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/models/deepseek_v2.py
View File

@@ -1731,12 +1731,12 @@ class DeepseekV2ForCausalLM(nn.Module):
disable_reason = None
if (
not _is_cuda
or torch.cuda.get_device_capability("cuda") < (9, 0)
or torch.cuda.get_device_capability("cuda") < (8, 0)
or self.config.architectures[0] != architecture
or self.config.n_routed_experts != 256
or self.config.n_shared_experts != 1
):
disable_reason = "Only Deepseek V3/R1 on NV-platform with capability >= 90 can use shared experts fusion optimization."
disable_reason = "Only Deepseek V3/R1 on NV-platform with capability >= 80 can use shared experts fusion optimization."
elif (
global_server_args_dict["enable_deepep_moe"]
or global_server_args_dict["enable_ep_moe"]
@@ -2040,7 +2040,7 @@ class DeepseekV2ForCausalLM(nn.Module):
if self.num_fused_shared_experts > 0:
assert self.num_fused_shared_experts == 1
logger.info("Shared experts fusion optimization enabled.")
log_info_on_rank0(logger, "Shared experts fusion optimization enabled.")
params_dict = dict(self.named_parameters())
weight_names = []