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

fix(dense): fix Qwen3.5 dense model precision bug in TP_SIZE>1 (#19070)

Browse Source
This commit is contained in:
Zheng Li
2026-02-25 20:54:42 +08:00
committed by GitHub
parent cdc411160b
commit d38c0e537d
1 changed files with 32 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
python/sglang/srt/models/qwen3_5.py
View File

@@ -400,11 +400,24 @@ class Qwen3_5LinearDecoderLayer(nn.Module):
use_reduce_scatter = self.layer_communicator.should_use_reduce_scatter(
forward_batch
)
hidden_states = self.mlp(hidden_states, forward_batch, use_reduce_scatter)
hidden_states, residual = self.layer_communicator.postprocess_layer(
hidden_states, residual, forward_batch
should_allreduce_fusion = (
self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer(
forward_batch
)
)
if isinstance(self.mlp, Qwen2MoeSparseMoeBlock):
hidden_states = self.mlp(hidden_states, forward_batch, use_reduce_scatter)
else:
hidden_states = self.mlp(
hidden_states, should_allreduce_fusion, use_reduce_scatter
)
if should_allreduce_fusion:
hidden_states._sglang_needs_allreduce_fusion = True
else:
hidden_states, residual = self.layer_communicator.postprocess_layer(
hidden_states, residual, forward_batch
)
return hidden_states, residual
@@ -633,11 +646,24 @@ class Qwen3_5AttentionDecoderLayer(nn.Module):
use_reduce_scatter = self.layer_communicator.should_use_reduce_scatter(
forward_batch
)
hidden_states = self.mlp(hidden_states, forward_batch, use_reduce_scatter)
hidden_states, residual = self.layer_communicator.postprocess_layer(
hidden_states, residual, forward_batch
should_allreduce_fusion = (
self.layer_communicator.should_fuse_mlp_allreduce_with_next_layer(
forward_batch
)
)
if isinstance(self.mlp, Qwen2MoeSparseMoeBlock):
hidden_states = self.mlp(hidden_states, forward_batch, use_reduce_scatter)
else:
hidden_states = self.mlp(
hidden_states, should_allreduce_fusion, use_reduce_scatter
)
if should_allreduce_fusion:
hidden_states._sglang_needs_allreduce_fusion = True
else:
hidden_states, residual = self.layer_communicator.postprocess_layer(
hidden_states, residual, forward_batch
)
return hidden_states, residual