[XPU] [CT] Enable CT W4A4MxFp4 path and add xpu kernel (#38896)

Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
Signed-off-by: zofia <110436990+zufangzhu@users.noreply.github.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/model_executor/kernels/linear/__init__.py

@@ -68,6 +68,9 @@ from vllm.model_executor.kernels.linear.mxfp4.flashinfer import (
 from vllm.model_executor.kernels.linear.mxfp4.marlin import (
     MarlinMxFp4LinearKernel,
 )
+from vllm.model_executor.kernels.linear.mxfp4.xpu import (
+    XPUMxFp4LinearKernel,
+)
 from vllm.model_executor.kernels.linear.mxfp8 import (
     Mxfp8LinearKernel,
     Mxfp8LinearLayerConfig,
@@ -291,6 +294,9 @@ _POSSIBLE_MXFP4_KERNELS: dict[PlatformEnum, list[type[MxFp4LinearKernel]]] = {
         FlashInferMxFp4LinearKernel,
         MarlinMxFp4LinearKernel,
     ],
+    PlatformEnum.XPU: [
+        XPUMxFp4LinearKernel,
+    ],
 }
 
 # TODO make all kernels inherit from MMLinearKernel

vllm/model_executor/kernels/linear/mxfp4/xpu.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
+    xpu_mxfp4_quantize as quant_mxfp4,
+)
+from vllm.model_executor.utils import replace_parameter
+from vllm.platforms import current_platform
+
+from .base import MxFp4LinearKernel, MxFp4LinearLayerConfig
+
+
+class XPUMxFp4LinearKernel(MxFp4LinearKernel):
+    """MXFP4 W4A4 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, "XPUMxFp4 only support on XPU"
+        return True, None
+
+    @classmethod
+    def can_implement(cls, c: MxFp4LinearLayerConfig) -> tuple[bool, str | None]:
+        return True, None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        weight = layer.weight.view(torch.float4_e2m1fn_x2)
+        replace_parameter(layer, "weight", weight.data.t())
+
+        weight_scale = layer.weight_scale.view(torch.float8_e8m0fnu)
+        weight_scale = weight_scale.t().contiguous()
+        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_fp4, x_blockscale = quant_mxfp4(x)
+        return torch.ops._xpu_C.fp4_gemm(
+            x_fp4,
+            layer.weight,
+            x_blockscale,
+            layer.weight_scale,
+            out_dtype,
+            bias,
+        )