biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][Model] Fix Qwen3.5/Qwen3Next ignoring --dtype flag on older GPUs (#35617)

Browse Source
This commit is contained in:
lailoo
2026-03-01 11:27:37 +08:00
committed by GitHub
parent 3ecd0bf9fc
commit afd089f231
2 changed files with 0 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/models/qwen3_5.py
View File

@@ -274,7 +274,6 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
1,
1,
config.hidden_size,
dtype=config.dtype,
),
)
self.ffn_layer_scale = torch.nn.Parameter(
@@ -282,7 +281,6 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
1,
1,
config.hidden_size,
dtype=config.dtype,
),
)

3
vllm/model_executor/models/qwen3_next.py
View File

@@ -463,7 +463,6 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
group_size=None,
norm_before_gate=True,
device=current_platform.current_device(),
dtype=config.dtype,
)
self.out_proj = RowParallelLinear(
@@ -1018,7 +1017,6 @@ class Qwen3NextDecoderLayer(nn.Module):
1,
1,
config.hidden_size,
dtype=config.dtype,
),
)
self.ffn_layer_scale = torch.nn.Parameter(
@@ -1026,7 +1024,6 @@ class Qwen3NextDecoderLayer(nn.Module):
1,
1,
config.hidden_size,
dtype=config.dtype,
),
)