biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] Always use forward_mulmat for Conv3d on newer versions of torch. (#38487)

Browse Source
This commit is contained in:
Roger Wang
2026-03-29 22:39:41 -07:00
committed by GitHub
parent bea23536f6
commit 92f0db57a8
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/layers/conv.py
View File

@@ -10,7 +10,7 @@ import torch.nn as nn
import torch.nn.functional as F
from vllm.model_executor.custom_op import CustomOp
from vllm.utils.torch_utils import is_torch_equal
from vllm.utils.torch_utils import is_torch_equal_or_newer
class ConvLayerBase(CustomOp):
@@ -252,11 +252,12 @@ class Conv3dLayer(ConvLayerBase):
return self._forward_conv(x)
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
# PyTorch2.9.0 disabled CUDNN's Conv3D, which caused a
# PyTorch 2.9.0+ disabled CUDNN's Conv3D, which caused a
# significant performance regression.
# See: https://github.com/vllm-project/vllm/issues/27406
# and https://github.com/pytorch/pytorch/issues/166122
# and https://github.com/huggingface/transformers/pull/45041
# By default, we use CUDNN's convolution ops with optimization.
if self.enable_linear and (is_torch_equal("2.9.0") or is_torch_equal("2.9.1")):
if self.enable_linear and is_torch_equal_or_newer("2.9.0"):
return self._forward_mulmat(x)
return self._forward_conv(x)