Support compress-tensors with nvfp4 or fp8 weights and modelopt with nvfp4 weights on Turing (#33076)

Signed-off-by: IriKa Qiu <qiujie.jq@gmail.com>
This commit is contained in:
IriKa
2026-01-27 11:04:05 -05:00
committed by GitHub
parent 0cd259b2d8
commit 66e601ef79
3 changed files with 5 additions and 4 deletions
@@ -96,6 +96,7 @@ from vllm.model_executor.parameter import (
PerTensorScaleParameter,
)
from vllm.model_executor.utils import replace_parameter
from vllm.platforms import current_platform
from vllm.utils.flashinfer import (
flashinfer_scaled_fp4_mm,
has_flashinfer,
@@ -1110,7 +1111,7 @@ class ModelOptNvFp4LinearMethod(LinearMethodBase):
self.backend = "none"
if envs.VLLM_NVFP4_GEMM_BACKEND is None:
if has_flashinfer():
if current_platform.has_device_capability(100) and has_flashinfer():
self.backend = "flashinfer-cutlass"
elif cutlass_fp4_supported():
self.backend = "cutlass"