[XPU] support MLA model on Intel GPU (#37143)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/model_executor/layers/attention/mla_attention.py

@@ -1059,6 +1059,10 @@ except ImportError:
                 "MLA models using TRITON_MLA will require flash_attn. "
                 "AITER_MLA backends use aiter kernels instead."
             )
+    elif current_platform.is_xpu():
+        from vllm._xpu_ops import xpu_ops as ops
+
+        flash_attn_varlen_func = ops.flash_attn_varlen_func  # type: ignore[no-redef]
 
 
 def dynamic_per_batched_tensor_quant(

vllm/model_executor/layers/quantization/input_quant_fp8.py

@@ -165,6 +165,16 @@ class QuantFP8(CustomOp):
         # Fallback to CUDA implementation
         return self.forward_cuda(x, scale, scale_ub)
 
+    def forward_xpu(
+        self,
+        x: torch.Tensor,
+        scale: torch.Tensor | None = None,
+        scale_ub: torch.Tensor | None = None,
+        use_triton: bool = False,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        # XPU can use same code path as CUDA.
+        return self.forward_cuda(x, scale, scale_ub, use_triton)
+
     def forward_native(
         self,
         x: torch.Tensor,