From ab7c4ab6bb83a05461387ce6ab26500f31d6bf61 Mon Sep 17 00:00:00 2001 From: Duyi-Wang Date: Wed, 3 Jun 2026 14:01:19 +0800 Subject: [PATCH] [AMD] Fix correctness for AITER MLA backend with `--page-size > 1` (#25556) --- python/sglang/srt/layers/attention/aiter_backend.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py index 61f0c20aa..982b98677 100755 --- a/python/sglang/srt/layers/attention/aiter_backend.py +++ b/python/sglang/srt/layers/attention/aiter_backend.py @@ -488,8 +488,8 @@ class AiterAttnBackend(AttentionBackend): tile_q = 256 qhead_granularity = gqa_ratio qlen_granularity = tile_q // qhead_granularity - kvlen_granularity = max(128, self.page_size) - block_size = self.page_size + kvlen_granularity = 128 + block_size = 1 qo_indptr_cpu = qo_indptr.to("cpu", dtype=torch.int32) kv_indptr_cpu = kv_indptr.to("cpu", dtype=torch.int32)