[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
@@ -3,14 +3,18 @@
import torch
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization.utils.humming_utils import (
convert_linear_layer_to_humming_standard,
prepare_humming_layer,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import kMxfp4Dynamic
from vllm.platforms import current_platform
from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
logger = init_logger(__name__)
class HummingMxFp4LinearKernel(MxFp4LinearKernel):
"""Humming GEMM Kernel for MXFP4."""
@@ -28,7 +32,15 @@ class HummingMxFp4LinearKernel(MxFp4LinearKernel):
return True, None
@classmethod
def can_implement(cls, c: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
def can_implement(cls, config: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
if config.activation_quant_key not in (None, kMxfp4Dynamic):
return False, "only supports MXFP4 dynamic or unquantized activations"
if config.activation_quant_key is not None:
logger.warning_once(
"HummingMxFp4LinearKernel is a weight-only (A16) kernel; "
"the requested activation quantization (%s) is ignored.",
config.activation_quant_key,
)
return True, None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None: