xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-21 20:05:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix Qwen3-VL-MoE weight loading for EP (#25300)

Browse Source 
Signed-off-by: Roger Wang <hey@rogerw.io>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Roger Wang 2025-09-20 00:04:05 -07:00 committed by yewentao256
parent 267b4421b7
commit 0ac65d171b
1 changed files with 7 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/models/qwen3_vl_moe.py
View File

@ -122,9 +122,10 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
def load_fused_expert_weights(self, name: str, params_dict: dict, def load_fused_expert_weights(self, name: str, params_dict: dict,
loaded_weight: torch.Tensor, shard_id: str, loaded_weight: torch.Tensor, shard_id: str,
num_experts: int): num_experts: int) -> bool:
param = params_dict[name] param = params_dict[name]
weight_loader = typing.cast(Callable[..., bool], param.weight_loader) weight_loader = typing.cast(Callable[..., bool], param.weight_loader)
loaded_local_expert = False
for expert_id in range(num_experts): for expert_id in range(num_experts):
curr_expert_weight = loaded_weight[expert_id] curr_expert_weight = loaded_weight[expert_id]
success = weight_loader(param, success = weight_loader(param,
@ -133,9 +134,10 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
shard_id, shard_id,
expert_id, expert_id,
return_success=True) return_success=True)
if not success: if success:
return False loaded_local_expert = True
return True
return loaded_local_expert
def load_weights(self, weights: Iterable[tuple[str, def load_weights(self, weights: Iterable[tuple[str,
torch.Tensor]]) -> set[str]: torch.Tensor]]) -> set[str]: