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

enable DeepSeek V3 shared_experts_fusion in sm90 (#5571)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-04-21 03:46:42 +08:00
committed by GitHub
parent 0a0dd34e6a
commit d9dd529854
1 changed files with 12 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
python/sglang/srt/models/deepseek_v2.py
View File

@@ -1427,6 +1427,18 @@ class DeepseekV2ForCausalLM(nn.Module):
assert ( assert (
self.n_share_experts_fusion == self.tp_size self.n_share_experts_fusion == self.tp_size
), f"Shared experts fusion optimization is enabled in DeepSeek V3/R1, set it to {self.tp_size} can get best optimized performace." ), 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)
and self.config.architectures[0] == "DeepseekV3ForCausalLM"
and self.config.n_routed_experts == 256
and (not global_server_args_dict["enable_deepep_moe"])
):
self.n_share_experts_fusion = self.tp_size
global_server_args_dict["n_share_experts_fusion"] = self.tp_size
logger.info(
"Deepseek V3/R1 with fp8 can use shared experts fusion optimization when SM version >=90. Shared experts fusion optimization is enabled."
)
self.model = DeepseekV2Model( self.model = DeepseekV2Model(
config, quant_config, prefix=add_prefix("model", prefix) config, quant_config, prefix=add_prefix("model", prefix)