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, 14:45 UTC
[Bugfix] Only use triton_kernels for MXFP4 on SM90 and SM100 (#29339)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Nov 24, 2025 at 18:22 UTC
c17610e2baf5e40b3b0638b272bfe7e04e471bfe
1 file changed
+9
-6
vllm/model_executor/layers/quantization/mxfp4.py
+9
-6
@@ -132,12 +132,15 @@ def get_mxfp4_backend(with_lora_support: bool) -> Mxfp4Backend:
132
)
133
134
# If FlashInfer is not available, try either Marlin or Triton
135
- if (
136
- envs.VLLM_MXFP4_USE_MARLIN
137
- or current_platform.get_device_capability()[0] < 9
138
- or not has_triton_kernels()
139
- or not is_torch_equal_or_newer("2.8.0")
140
- ):
135
+ triton_kernels_supported = (
136
+ has_triton_kernels()
137
+ and is_torch_equal_or_newer("2.8.0")
138
+ # NOTE: triton_kernels are only confirmed to work on SM90 and SM100
139
+ # SM110 fails with this error: https://github.com/vllm-project/vllm/issues/29317
140
+ # SM120 needs this fix: https://github.com/triton-lang/triton/pull/8498
141
+ and (9, 0) <= current_platform.get_device_capability() < (11, 0)
142
+ )
143
+ if envs.VLLM_MXFP4_USE_MARLIN or not triton_kernels_supported:
144
logger.info_once("Using Marlin backend")
145
return Mxfp4Backend.MARLIN
146
else: