[NPU][BugFix] Avoid M-RoPE recompilation for variable sequence lengths (#40371)

Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-09-22 08:19:25 +08:00
parent 61d0cf2074
commit c53cc8e1eb
1 changed files with 5 additions and 2 deletions
@@ -7,11 +7,14 @@ import triton
 import triton.language as tl


-@triton.jit
+# Sequence length varies across requests. Specializing S would create one
+# compiled kernel variant per observed length, while the existing mask handles
+# the final partial block.
+@triton.jit(do_not_specialize=["S"])
 def apply_interleaved_rope_kernel(
    x_ptr,
    out_ptr,
-    S: tl.constexpr,
+    S,
    D: tl.constexpr,
    stride_x_m,
    stride_x_s,