[XPU][6/N] add xpu scaled_mm kernel (#34117)

Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
This commit is contained in:
zofia
2026-02-09 20:17:35 +08:00
committed by GitHub
parent caad9f1e01
commit 9bdb06b436
4 changed files with 67 additions and 10 deletions
@@ -39,6 +39,9 @@ from vllm.model_executor.layers.quantization.kernels.scaled_mm.ScaledMMLinearKer
from vllm.model_executor.layers.quantization.kernels.scaled_mm.triton import (
TritonInt8ScaledMMLinearKernel,
)
from vllm.model_executor.layers.quantization.kernels.scaled_mm.xpu import (
XPUFP8ScaledMMLinearKernel,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import QuantKey
from vllm.platforms import PlatformEnum, current_platform
@@ -72,6 +75,9 @@ _POSSIBLE_FP8_KERNELS: dict[PlatformEnum, list[type[FP8ScaledMMLinearKernel]]] =
PerTensorTorchFP8ScaledMMLinearKernel,
ChannelWiseTorchFP8ScaledMMLinearKernel,
],
PlatformEnum.XPU: [
XPUFP8ScaledMMLinearKernel,
],
}
_KernelT = TypeVar("_KernelT", bound=ScaledMMLinearKernel)
@@ -0,0 +1,59 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Sequence
import torch
from vllm.model_executor.layers.quantization.kernels.scaled_mm.ScaledMMLinearKernel import ( # noqa: E501
FP8ScaledMMLinearKernel,
FP8ScaledMMLinearLayerConfig,
)
from vllm.platforms import current_platform
class XPUFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
@classmethod
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if not current_platform.is_xpu():
return False, "XPUFP8ScaledMM only support on XPU"
return True, None
@classmethod
def can_implement(cls, c: FP8ScaledMMLinearLayerConfig) -> tuple[bool, str | None]:
if c.weight_quant_key.dtype not in {torch.float8_e5m2, torch.float8_e4m3fn}:
return False, "XPUFP8ScaledMM only support FP8 weight dtype"
return True, None
def __init__(
self, c: FP8ScaledMMLinearLayerConfig, layer_param_names: Sequence[str]
) -> None:
assert self.can_implement(c)[0]
assert self.is_supported()[0]
self.config = c
self.layer_param_names = layer_param_names
def apply_weights(
self,
layer: torch.nn.Module,
x: torch.Tensor,
bias: torch.Tensor | None = None,
) -> torch.Tensor:
weight = layer.weight
weight_scale = layer.weight_scale
return torch.ops._xpu_C.fp8_gemm_w8a16(x, weight, weight_scale, bias)
def apply_scaled_mm(
self,
*,
A: torch.Tensor,
B: torch.Tensor,
out_dtype: torch.dtype,
As: torch.Tensor,
Bs: torch.Tensor,
bias: torch.Tensor | None,
output_shape: list,
) -> torch.Tensor:
pass