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 11, 05:00 UTC
[LoRA] Set default MXFP4 LoRA backend to Marlin (#30598)
Signed-off-by: Xin Yang <xyangx@amazon.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Xin Yang committed
Dec 18, 2025 at 08:42 UTC
9a5e96523be33f7ddd5aa56421b1e41000c0f2e2
1 file changed
+5
-5
vllm/model_executor/layers/quantization/mxfp4.py
+5
-5
@@ -95,12 +95,12 @@ def get_mxfp4_backend_with_lora() -> Mxfp4Backend:
95
# SM120 needs this fix: https://github.com/triton-lang/triton/pull/8498
96
and (9, 0) <= current_platform.get_device_capability() < (11, 0)
97
)
98
- if envs.VLLM_MXFP4_USE_MARLIN or not triton_kernels_supported:
99
- logger.info_once("[get_mxfp4_backend_with_lora] Using Marlin backend")
100
- return Mxfp4Backend.MARLIN
98
+ if envs.VLLM_MXFP4_USE_MARLIN is False and triton_kernels_supported:
99
+ logger.info_once("[get_mxfp4_backend_with_lora] Using Triton backend")
100
+ return Mxfp4Backend.TRITON
101
102
- logger.info_once("[get_mxfp4_backend_with_lora] Using Triton backend")
103
- return Mxfp4Backend.TRITON
102
+ logger.info_once("[get_mxfp4_backend_with_lora] Using Marlin backend")
103
+ return Mxfp4Backend.MARLIN
104
105
106
def get_mxfp4_backend(with_lora_support: bool) -> Mxfp4Backend: