[1/2] Speed up prefill mla attention (#10156)

sgl-kernel/python/sgl_kernel/__init__.py

@@ -23,6 +23,7 @@ from sgl_kernel.cutlass_moe import cutlass_w4a8_moe_mm, get_cutlass_w4a8_moe_mm_
 from sgl_kernel.elementwise import (
     FusedSetKVBufferArg,
     apply_rope_with_cos_sin_cache_inplace,
+    concat_mla_k,
     copy_to_gpu_no_ce,
     downcast_fp8,
     fused_add_rmsnorm,

sgl-kernel/python/sgl_kernel/elementwise.py

@@ -371,3 +371,11 @@ def downcast_fp8(
 
 def copy_to_gpu_no_ce(input: List[int], output: torch.Tensor):
     torch.ops.sgl_kernel.copy_to_gpu_no_ce(input, output)
+
+
+def concat_mla_k(
+    k: torch.Tensor,
+    k_nope: torch.Tensor,
+    k_rope: torch.Tensor,
+):
+    torch.ops.sgl_kernel.concat_mla_k(k, k_nope, k_rope)