[AMD] fix performance regression issue when run gpt-oss with "--context-length 13824" (#21691)

2026-03-30 16:30:16 -07:00
parent 4b8456e266
commit 67c295b5f5
1 changed files with 1 additions and 1 deletions
@@ -2503,7 +2503,7 @@ class AiterAttnBackend(AttentionBackend):

                o = torch.empty_like(q, dtype=self.input_dtype)

-                max_kv_len = page_table.shape[1]
+                max_kv_len = page_table.shape[1] * self.page_size

                unified_attention(
                    q=q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),