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 14, 05:30 UTC
[Bugfix] fix Qwen3VLMoe load when pp > 1 (#25838)
Signed-off-by: liuye.hj <liuye.hj@alibaba-inc.com> Co-authored-by: liuye.hj <liuye.hj@alibaba-inc.com>
JJJYmmm committed
Sep 29, 2025 at 01:56 UTC
471997adf696f4f97b7b6bc7b52610e1eceac54c
1 file changed
+2
-2
vllm/model_executor/models/qwen3_vl_moe.py
+2
-2
@@ -212,6 +212,8 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
212
# attempted to load as other weights later
213
is_expert_weight = True
214
name_mapped = name.replace(weight_name, param_name)
215
+ if is_pp_missing_parameter(name_mapped, self):
216
+ continue
217
if is_fused_expert:
218
loaded_weight = loaded_weight.transpose(-1,
219
-2) # no bias
@@ -230,8 +232,6 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
232
name_mapped, params_dict, loaded_weight,
233
shard_id, num_experts)
234
else:
233
- if is_pp_missing_parameter(name_mapped, self):
234
- continue
235
# Skip loading extra parameters for GPTQ/modelopt models
236
if name_mapped.endswith(
237
ignore_suffixes