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

fix dual stream bug (#10352)

Browse Source
This commit is contained in:
Yi Zhang
2025-09-12 11:53:42 +08:00
committed by GitHub
parent 46d8fb1c98
commit 27778010fc
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/models/qwen2_moe.py
View File

@@ -62,6 +62,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from sglang.srt.managers.schedule_batch import global_server_args_dict
from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.two_batch_overlap import model_forward_maybe_tbo
@@ -194,7 +195,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
) -> torch.Tensor:
current_stream = torch.cuda.current_stream()
self.alt_stream.wait_stream(current_stream)
shared_output = self._forward_shared_experts(hidden_states)
shared_output = self._forward_shared_experts(hidden_states.clone())
with torch.cuda.stream(self.alt_stream):
router_output = self._forward_router_experts(hidden_states)
@@ -217,6 +218,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
self.alt_stream is not None
and hidden_states.shape[0] > 0
and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
and get_is_capture_mode()
):
final_hidden_states, shared_output = self.forward_normal_dual_stream(
hidden_states