class HummingMxFp4LinearKernel(MxFp4LinearKernel):
"""Humming GEMM Kernel for MXFP4."""
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if not current_platform.is_cuda():
return False, "Humming only supported on CUDA"
if not has_humming():
return False, "Humming is not installed"
if not current_platform.has_device_capability(75):
return False, "Humming only supported on SM75+"
return True, None
@classmethod
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:
layer.weight_scale.data = layer.weight_scale.data.view(torch.float8_e8m0fnu)
name_map = {"weight": "weight", "weight_scale": "weight_scale"}
quant_config = {
"quant_method": "humming",
"dtype": "float4e2m1",
"scale_dtype": "float8e8m0",
"group_size": 32,
"weight_scale_type": "group",
}
convert_linear_layer_to_humming_standard(layer=layer, name_map=name_map)
self.layer_config = prepare_humming_linear_layer_config(layer, quant_config)
self.compute_config = get_humming_linear_compute_config()
self.locks = torch.zeros(1024, dtype=torch.int32, device=layer.weight.device)
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
return apply_humming_linear(
layer,
x,
layer_config=self.layer_config,
compute_config=self.compute_config,
locks=self.locks,
)