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][CPU] Skip unsupported custom op register on CPU (#25534)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Sep 24, 2025 at 14:21 UTC
1cbcfb94ded21da2b922b452c1e90baef87e419b
1 file changed
+8
-6
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+8
-6
@@ -130,12 +130,14 @@ def _w8a8_triton_block_scaled_mm_fake(
130
device=qx.device)
131
132
133
-direct_register_custom_op(
134
- "w8a8_triton_block_scaled_mm_func",
135
- _w8a8_triton_block_scaled_mm_func,
136
- fake_impl=_w8a8_triton_block_scaled_mm_fake,
137
- dispatch_key="CUDA",
138
-)
133
+# Note: the check can be removed when CPU torch > 2.7
134
+if not current_platform.is_cpu():
135
+ direct_register_custom_op(
136
+ "w8a8_triton_block_scaled_mm_func",
137
+ _w8a8_triton_block_scaled_mm_func,
138
+ fake_impl=_w8a8_triton_block_scaled_mm_fake,
139
+ dispatch_key="CUDA",
140
+ )
141
142
143
# TODO fix ROCm->Triton custom path: