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

2025-09-09 00:00:33 +08:00
parent 2c2b19b18b
commit 0096798ed6
6 changed files with 130 additions and 0 deletions
--- a/sgl-kernel/python/sgl_kernel/elementwise.py
+++ b/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)