Refactor CT NVFP4 linear to use a single class (#42443)
This commit is contained in:
@@ -842,7 +842,7 @@ _NVFP4_BACKEND_TO_KERNEL: dict[str, type[NvFp4LinearKernel]] = {
|
||||
}
|
||||
|
||||
|
||||
def init_nvfp4_linear_kernel() -> NvFp4LinearKernel:
|
||||
def init_nvfp4_linear_kernel(use_a16: bool = False) -> NvFp4LinearKernel:
|
||||
"""Select and instantiate the best NVFP4 linear kernel for the
|
||||
current platform."""
|
||||
config = NvFp4LinearLayerConfig()
|
||||
@@ -885,7 +885,9 @@ def init_nvfp4_linear_kernel() -> NvFp4LinearKernel:
|
||||
elif linear_backend == "auto":
|
||||
# Deprecated env-var overrides — only honoured when --linear-backend
|
||||
# is "auto". Deprecation warnings are emitted from vllm/envs.py.
|
||||
if envs.VLLM_USE_FBGEMM:
|
||||
if use_a16: # force a16 if running weight-only quantization
|
||||
force_kernel = MarlinNvFp4LinearKernel
|
||||
elif envs.VLLM_USE_FBGEMM:
|
||||
force_kernel = FbgemmNvFp4LinearKernel
|
||||
elif envs.VLLM_USE_NVFP4_CT_EMULATIONS:
|
||||
force_kernel = EmulationNvFp4LinearKernel
|
||||
|
||||
Reference in New Issue
Block a user