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 12, 19:45 UTC
[BugFix] skip combo kernel on cpu (#29129)
Signed-off-by: Boyuan Feng <boyuan@meta.com>
Boyuan Feng committed
Nov 20, 2025 at 19:50 UTC
8c25f9cfb619c46ecaf8fdb1ef697da9cc21d7e0
1 file changed
+2
vllm/config/compilation.py
+2
@@ -664,6 +664,8 @@ class CompilationConfig:
664
is_torch_equal_or_newer("2.9.0.dev")
665
and "combo_kernels" not in self.inductor_compile_config
666
and "benchmark_combo_kernel" not in self.inductor_compile_config
667
+ # (fixme @boyuan) combo kernel does not support cpu yet.
668
+ and not current_platform.is_cpu()
669
):
670
# use horizontal fusion, which is useful for fusing qk-norm and
671
# qk-rope when query and key have different shapes.