A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 7, 12:45 UTC
[Bugfix][Model] Fix Qwen3.5/Qwen3Next ignoring --dtype flag on older GPUs (#35617)
lailoo committed
Mar 1, 2026 at 11:27 UTC
afd089f231d714e7fd06b51e3bc7df7fe004c7f9
2 files changed
-5
vllm/model_executor/models/qwen3_5.py
-2
@@ -274,7 +274,6 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
274
1,
275
1,
276
config.hidden_size,
277
- dtype=config.dtype,
277
),
278
)
279
self.ffn_layer_scale = torch.nn.Parameter(
@@ -282,7 +281,6 @@ class Qwen3_5DecoderLayer(Qwen3NextDecoderLayer):
281
1,
282
1,
283
config.hidden_size,
285
- dtype=config.dtype,
284
),
285
)
286
vllm/model_executor/models/qwen3_next.py
-3
@@ -463,7 +463,6 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
463
group_size=None,
464
norm_before_gate=True,
465
device=current_platform.current_device(),
466
- dtype=config.dtype,
466
)
467
468
self.out_proj = RowParallelLinear(
@@ -1018,7 +1017,6 @@ class Qwen3NextDecoderLayer(nn.Module):
1017
1,
1018
1,
1019
config.hidden_size,
1021
- dtype=config.dtype,
1020
),
1021
)
1022
self.ffn_layer_scale = torch.nn.Parameter(
@@ -1026,7 +1024,6 @@ class Qwen3NextDecoderLayer(nn.Module):
1024
1,
1025
1,
1026
config.hidden_size,
1029
- dtype=config.dtype,
1027
),
1028
)
1029