[XPU] Use SYCL kernels for topk_transform on XPU (#33318)

Signed-off-by: Cui, Lily <lily.cui@intel.com>
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

python/sglang/kernels/ops/attention/dsv4/topk.py

@@ -11,6 +11,7 @@ from sglang.kernels.jit.utils import (
     load_jit,
     make_cpp_args,
 )
+from sglang.srt.utils import is_xpu
 
 from .utils import make_name
 
@@ -58,6 +59,10 @@ def topk_transform_512(
         torch.ops.sgl_kernel.deepseek_v4_topk_transform_512(
             scores, seq_lens, page_tables, out_page_indices, page_size, out_raw_indices
         )
+    elif is_xpu():
+        torch.ops.sgl_kernel.topk_transform(
+            scores, seq_lens, page_tables, out_page_indices, page_size, out_raw_indices
+        )
     else:
         module = _jit_topk_v1_module()
         module.topk_transform(
@@ -111,6 +116,15 @@ def topk_transform_ragged_v2(
     They are invalid for that row and the buffer must have no other consumer.
     ``seq_lens`` entries must be NON-NEGATIVE, as for the paged entry point.
     """
+    if is_xpu():
+        torch.ops.sgl_kernel.topk_transform_ragged(
+            scores,
+            seq_lens,
+            out_indices,
+            out_offsets,
+            row_starts,
+        )
+        return
     module = _jit_topk_v2_module()
     module.topk_transform_ragged(scores, seq_lens, row_starts, out_offsets, out_indices)
 
@@ -143,6 +157,16 @@ def topk_transform_512_v2(
     the valid way to express "no tokens": the row takes the trivial path and
     the output is all -1.
     """
+    if is_xpu():
+        torch.ops.sgl_kernel.topk_transform_paged(
+            scores,
+            seq_lens,
+            page_tables,
+            out_page_indices,
+            page_size,
+            metadata,
+        )
+        return
     module = _jit_topk_v2_module()
     module.topk_transform_paged(
         scores,

python/sglang/srt/layers/attention/dsv4/metadata.py

@@ -152,9 +152,9 @@ class PagedIndexerMetadata:
 
             assert isinstance(self.deep_gemm_metadata, torch.Tensor)
 
-        from sglang.kernels.ops.attention.dsv4 import plan_topk_v2
+        if envs.SGLANG_OPT_USE_TOPK_V2.get() and not is_xpu():
+            from sglang.kernels.ops.attention.dsv4 import plan_topk_v2
 
-        if envs.SGLANG_OPT_USE_TOPK_V2.get():
             self.topk_metadata = plan_topk_v2(self.c4_seq_lens)
         else:
             self.topk_metadata = torch.empty((0,))