[XPU][Diffusion] Enable MiniMax H3 on XPU platforms (#33366)

python/sglang/multimodal_gen/runtime/layers/attention/backends/xpu_backend.py

@@ -119,3 +119,30 @@ class XPUAttentionImpl(AttentionImpl):
 
         result = out.reshape(bsz, seqlen_q, nheads_q, d)
         return result
+
+    def forward_varlen(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        *,
+        cu_seqlens: torch.Tensor,
+        max_seqlen: int,
+        cu_seqlens_host: tuple[int, ...] | None = None,
+    ) -> torch.Tensor:
+        del cu_seqlens_host
+        q_ = query.contiguous()
+        k_ = key.contiguous()
+        v_ = value.contiguous()
+        output = flash_attn_func(
+            q=q_,
+            k=k_,
+            v=v_,
+            cu_seqlens_q=cu_seqlens,
+            cu_seqlens_k=cu_seqlens,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            softmax_scale=self.softmax_scale,
+            causal=self.causal,
+        )
+        return output[0] if isinstance(output, tuple) else output

python/sglang/multimodal_gen/runtime/managers/memory_managers/component_manager.py

@@ -926,6 +926,7 @@ class ComponentResidencyManager:
             current_platform.is_cuda()
             or current_platform.is_rocm()
             or current_platform.is_npu()
+            or current_platform.is_xpu()
         )
         return is_supported_platform and current_platform.is_device_type(
             self._module_device(module)

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/keyframe_encoding.py

@@ -48,6 +48,7 @@ def minimax_h3_scoped_encode_rng(seed: int, device: torch.device | None = None):
         current_platform.is_cuda()
         or current_platform.is_rocm()
         or current_platform.is_npu()
+        or current_platform.is_xpu()
     )
     if (
         device is not None

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py

@@ -679,9 +679,10 @@ class MiniMaxH3DenoisingStage(DenoisingStage):
             or current_platform.is_cpu()
             or current_platform.is_mps()
             or current_platform.is_npu()
+            or current_platform.is_xpu()
         ):
             raise RuntimeError(
-                "MiniMax H3 full-loop denoise requires CPU, CUDA, MPS, or Ascend NPU"
+                "MiniMax H3 full-loop denoise requires CPU, CUDA, MPS, XPU, or Ascend NPU"
             )
 
         device = current_platform.get_local_torch_device()