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 _synced_weight_loader (#24565)
Signed-off-by: Kyuyeun Kim <kyuyeunk@google.com>
Kyuyeun Kim committed
Sep 11, 2025 at 01:52 UTC
ed5ae4aaceacc28461f7e324c8f26be6fadd59df
1 file changed
+2
-1
vllm/model_executor/utils.py
+2
-1
@@ -52,10 +52,11 @@ def set_weight_attrs(
52
def _make_synced_weight_loader(original_weight_loader):
53
54
def _synced_weight_loader(param, *args, **kwargs):
55
- original_weight_loader(param, *args, **kwargs)
55
+ out = original_weight_loader(param, *args, **kwargs)
56
# torch._sync doesn't support, is not needed for CPU tensors.
57
if param.device != torch.device("cpu"):
58
torch._sync(param)
59
+ return out
60
61
return _synced_weight_loader
62