[ROCm][Quark][7/N] Use MXFP4 linear kernel abstraction for emulation backend (#48949)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
fxmarty-amd
2026-07-31 10:07:15 -07:00
committed by GitHub
parent d87d2ca747
commit e67a2e0a56
17 changed files with 824 additions and 90 deletions
@@ -7,6 +7,9 @@ from torch.nn.parameter import Parameter
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
swizzle_mxfp4_scales,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
kMxfp4Dynamic,
)
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer_cutedsl
@@ -28,6 +31,8 @@ class FlashInferMxFp4LinearKernel(MxFp4LinearKernel):
@classmethod
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
if config.activation_quant_key != kMxfp4Dynamic:
return False, "only supports MXFP4 dynamic activation"
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None: