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
[CPU] Disable oneDNN linear on non-x86 platforms (#25166)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Sep 19, 2025 at 15:27 UTC
8c1d4acbfe70a1dffdb9a3db57b4d12329350295
1 file changed
+3
-2
vllm/model_executor/layers/utils.py
+3
-2
@@ -7,7 +7,7 @@ import torch
7
8
from vllm import _custom_ops as ops
9
from vllm import envs
10
-from vllm.platforms import current_platform
10
+from vllm.platforms import CpuArchEnum, current_platform
11
from vllm.utils import direct_register_custom_op
12
13
@@ -167,7 +167,8 @@ def dispatch_cpu_unquantized_gemm(
167
if remove_weight:
168
layer.weight = torch.nn.Parameter(torch.empty(0),
169
requires_grad=False)
170
- elif ops._supports_onednn:
170
+ elif (ops._supports_onednn
171
+ and current_platform.get_cpu_architecture() == CpuArchEnum.X86):
172
origin_weight = layer.weight
173
if remove_weight:
174
layer.weight = torch.nn.Parameter(torch.empty(0),