minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[DeepSeek V3] Run routed experts on main stream in dual-stream MoE (#29142)

Browse Source 
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Khoa Pham
2026-06-26 00:48:59 -07:00
committed by GitHub
co-authored by Cursor Claude Opus 4.8
parent aeb4e98108
commit 999199f9ff
1 changed files with 47 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_v2.py
+47 -45
View File
@@ -936,59 +936,61 @@ class DeepseekV2MoE(nn.Module):
) -> torch.Tensor: ) -> torch.Tensor:
current_stream = torch.cuda.current_stream() current_stream = torch.cuda.current_stream()
self.alt_stream.wait_stream(current_stream) self.alt_stream.wait_stream(current_stream)
shared_output = self._forward_shared_experts(
hidden_states, gemm_output_zero_allocator
)
server_args = get_global_server_args() server_args = get_global_server_args()
dispatch_info = ( dispatch_info = (
ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id) ExpertLocationDispatchInfo.init_new(layer_id=self.layer_id)
if server_args.enable_eplb if server_args.enable_eplb
else None else None
) )
# router_logits: (num_tokens, n_experts)
router_logits = self.gate(hidden_states, gemm_output_zero_allocator)
if use_flashinfer_trtllm_bypass:
topk_output = BypassedTopKOutput(
hidden_states=hidden_states,
router_logits=router_logits,
topk_config=self.topk.topk_config,
)
else:
topk_kwargs = (
{"input_ids": input_ids_global}
if getattr(self, "is_hash", False)
else {}
)
topk_output = self.topk(
hidden_states,
router_logits,
expert_location_dispatch_info=dispatch_info,
**topk_kwargs,
)
has_shared_output = (
hidden_states.shape[0] > 0 and self.num_fused_shared_experts == 0
)
deferred_finalize = (
has_shared_output
and not self._shared_expert_tp1
and topk_output.format == TopKOutputFormat.BYPASSED
and self.experts.supports_deferred_finalize
)
if deferred_finalize:
final_hidden_states = self.experts.forward_deferred_finalize(
hidden_states, topk_output
)
elif use_flashinfer_trtllm_bypass:
final_hidden_states = self.experts.forward_impl(hidden_states, topk_output)
else:
final_hidden_states = self.experts(hidden_states, topk_output)
if (
not _is_cuda
and not _is_musa
and not _use_aiter
or isinstance(self.experts.quant_method, KTEPWrapperMethod)
):
final_hidden_states *= self.routed_scaling_factor
with torch.cuda.stream(self.alt_stream): with torch.cuda.stream(self.alt_stream):
# router_logits: (num_tokens, n_experts) shared_output = self._forward_shared_experts(
router_logits = self.gate(hidden_states, gemm_output_zero_allocator) hidden_states, gemm_output_zero_allocator
if use_flashinfer_trtllm_bypass:
topk_output = BypassedTopKOutput(
hidden_states=hidden_states,
router_logits=router_logits,
topk_config=self.topk.topk_config,
)
else:
topk_kwargs = (
{"input_ids": input_ids_global}
if getattr(self, "is_hash", False)
else {}
)
topk_output = self.topk(
hidden_states,
router_logits,
expert_location_dispatch_info=dispatch_info,
**topk_kwargs,
)
deferred_finalize = (
shared_output is not None
and not self._shared_expert_tp1
and topk_output.format == TopKOutputFormat.BYPASSED
and self.experts.supports_deferred_finalize
) )
if deferred_finalize:
final_hidden_states = self.experts.forward_deferred_finalize(
hidden_states, topk_output
)
elif use_flashinfer_trtllm_bypass:
final_hidden_states = self.experts.forward_impl(
hidden_states, topk_output
)
else:
final_hidden_states = self.experts(hidden_states, topk_output)
if (
not _is_cuda
and not _is_musa
and not _use_aiter
or isinstance(self.experts.quant_method, KTEPWrapperMethod)
):
final_hidden_states *= self.routed_scaling_factor
current_stream.wait_stream(self.alt_stream) current_stream.wait_stream(self.alt_stream)