[XPU] [Linear] add torch as xpu linear backend (#49664)

Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
zofia
2026-08-03 14:30:21 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent 2755489a26
commit e481da9508
5 changed files with 15 additions and 8 deletions
@@ -363,6 +363,8 @@ _POSSIBLE_FP8_KERNELS: dict[PlatformEnum, list[type[FP8ScaledMMLinearKernel]]] =
PlatformEnum.XPU: [
XPUW8A16FP8LinearKernel,
XPUW8A8FP8LinearKernel,
PerTensorTorchFP8ScaledMMLinearKernel,
ChannelWiseTorchFP8ScaledMMLinearKernel,
],
}
@@ -32,8 +32,12 @@ class TorchFP8ScaledMMLinearKernel(FP8ScaledMMLinearKernel):
def is_supported(
cls, compute_capability: int | None = None
) -> tuple[bool, str | None]:
if not (current_platform.is_cuda_alike() or current_platform.is_cpu()):
return False, "requires ROCm, CUDA or CPU."
if not (
current_platform.is_cuda_alike()
or current_platform.is_cpu()
or current_platform.is_xpu()
):
return False, "requires ROCm, CUDA, CPU or XPU."
if compute_capability is not None and compute_capability < 89:
return False, "requires compute capability 89 and above."
@@ -811,12 +811,11 @@ class CompressedTensorsConfig(QuantizationConfig):
if act_quant_format:
if self._is_fp8_w8a8(weight_quant, input_quant):
if current_platform.is_xpu():
# On XPU, --linear-backend xpu opts into W8A8 FP8
# On XPU, --linear-backend xpu or torch opts into W8A8 FP8
# linear kernel; otherwise default to W8A16.
config = get_current_vllm_config_or_none()
is_fp8_w8a8_supported = (
config is not None
and config.kernel_config.linear_backend == "xpu"
is_fp8_w8a8_supported = config is not None and (
config.kernel_config.linear_backend in ("xpu", "torch")
)
else:
is_fp8_w8a8_supported = self._check_scheme_supported(