chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

remove unecessary dual stream token threshold from the rest of models (qwen moe, kimi linear, etc.) (#14337)

Browse Source
This commit is contained in:
b8zhong
2025-12-06 19:57:26 -08:00
committed by GitHub
parent 91c9c14c28
commit 6d5d76ad97
4 changed files with 0 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/models/bailing_moe.py
View File

@@ -349,11 +349,9 @@ class BailingMoESparseMoeBlock(nn.Module):
num_tokens, hidden_size = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_size)
DUAL_STREAM_TOKEN_THRESHOLD = 1024
if (
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(

2
python/sglang/srt/models/kimi_linear.py
View File

@@ -125,13 +125,11 @@ class KimiMoE(nn.Module):
hidden_states = hidden_states.view(-1, hidden_size)
shared_output = None
DUAL_STREAM_TOKEN_THRESHOLD = 1024
if (
self.alt_stream is not None
and self.num_shared_experts is not None
and hidden_states.shape[0] > 0
and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
and get_is_capture_mode()
):
current_stream = torch.cuda.current_stream()

2
python/sglang/srt/models/llada2.py
View File

@@ -349,11 +349,9 @@ class LLaDA2MoeSparseMoeBlock(nn.Module):
num_tokens, hidden_size = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_size)
DUAL_STREAM_TOKEN_THRESHOLD = 1024
if (
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(

2
python/sglang/srt/models/qwen2_moe.py
View File

@@ -304,11 +304,9 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
if get_moe_a2a_backend().is_deepep():
return self._forward_deepep(hidden_states, forward_batch)
DUAL_STREAM_TOKEN_THRESHOLD = 1024
if (
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(