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 3, 13:00 UTC
Fix Weight loading for Qwen3.5-MTP and Qwen3-VL using runai_streamer (#42716)
Signed-off-by: weizhoublue <weizhou.lan@daocloud.io>
weizhoublue committed
May 17, 2026 at 08:54 UTC
a94189295b8b9c1d952be438b49ed5793db59159
2 files changed
+4
-4
vllm/model_executor/models/qwen3_5_mtp.py
+2
-2
@@ -175,8 +175,8 @@ class Qwen3_5MultiTokenPredictor(nn.Module):
175
param,
176
curr_expert_weight,
177
name,
178
- shard_id,
179
- expert_id,
178
+ shard_id=shard_id,
179
+ expert_id=expert_id,
180
return_success=True,
181
)
182
if success:
vllm/model_executor/models/qwen3_vl_moe.py
+2
-2
@@ -152,8 +152,8 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
152
param,
153
curr_expert_weight,
154
name,
155
- shard_id,
156
- expert_id,
155
+ shard_id=shard_id,
156
+ expert_id=expert_id,
157
return_success=True,
158
)
159
if success: