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 5, 01:00 UTC
[Kernel] [Helion] Force disable HOP path due to performance regression (#40171)
Signed-off-by: Yanan Cao <gmagogsfm@gmail.com> Co-authored-by: Claude Sonnet 4 <noreply@anthropic.com>
Yanan Cao committed
Apr 17, 2026 at 14:36 UTC
5cdddddd4a03b0d1b6fa1940458e432b91457ae1
1 file changed
+3
-2
vllm/kernels/helion/register.py
+3
-2
@@ -53,14 +53,15 @@ if not has_helion():
53
)
54
55
import helion
56
-from helion._compat import requires_torch_version
56
from helion.autotuner.base_search import BaseAutotuner
57
from helion.runtime.config import Config
58
from helion.runtime.settings import default_autotuner_fn
59
60
# TODO(gmagogsfm): Remove CustomOp fallback path (_get_or_register_custom_op,
61
# vllm_helion_lib, direct_register_custom_op) once vLLM requires PyTorch >= 2.11.
63
-_HOP_AVAILABLE = requires_torch_version("2.11")
62
+# FIXME(gmagogsfm): Re-enable HOP path once performance regression is fixed.
63
+# _HOP_AVAILABLE = requires_torch_version("2.11")
64
+_HOP_AVAILABLE = False
65
66
if _HOP_AVAILABLE:
67
from helion._compat import supports_torch_compile_fusion