wehub-resource-sync
52 lines
1.6 KiB
Python
52 lines
1.6 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import torch
|
|
|
|
from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
|
|
xpu_mxfp8_quantize as quant_mxfp8,
|
|
)
|
|
from vllm.model_executor.utils import replace_parameter
|
|
from vllm.platforms import current_platform
|
|
|
|
from .Mxfp8LinearKernel import Mxfp8LinearKernel, Mxfp8LinearLayerConfig
|
|
|
|
|
|
class XPUMxFp8LinearKernel(Mxfp8LinearKernel):
|
|
"""MXFP8 W8A8 GEMM on XPU."""
|
|
|
|
@classmethod
|
|
def is_supported(
|
|
cls, compute_capability: int | None = None
|
|
) -> tuple[bool, str | None]:
|
|
if not current_platform.is_xpu():
|
|
return False, "XPUMxFp8 only support on XPU"
|
|
return True, None
|
|
|
|
@classmethod
|
|
def can_implement(cls, c: Mxfp8LinearLayerConfig) -> tuple[bool, str | None]:
|
|
return True, None
|
|
|
|
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
|
weight_scale = layer.weight_scale.view(torch.float8_e8m0fnu)
|
|
weight_scale = weight_scale.t().contiguous()
|
|
replace_parameter(layer, "weight", layer.weight.t())
|
|
replace_parameter(layer, "weight_scale", weight_scale.data)
|
|
|
|
def apply_weights(
|
|
self,
|
|
layer: torch.nn.Module,
|
|
x: torch.Tensor,
|
|
bias: torch.Tensor | None = None,
|
|
) -> torch.Tensor:
|
|
out_dtype = x.dtype
|
|
x_fp8, x_scale = quant_mxfp8(x)
|
|
return torch.ops._xpu_C.fp8_gemm(
|
|
x_fp8,
|
|
layer.weight,
|
|
out_dtype,
|
|
x_scale,
|
|
layer.weight_scale,
|
|
bias,
|
|
)
|