minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] Fix correctness for AITER MLA backend with --page-size > 1 (#25556)

Browse Source
This commit is contained in:
Duyi-Wang
2026-06-02 23:01:19 -07:00
committed by GitHub
parent 6d53615699
commit ab7c4ab6bb
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/aiter_backend.py
+2 -2
View File
@@ -488,8 +488,8 @@ class AiterAttnBackend(AttentionBackend):
tile_q = 256 tile_q = 256
qhead_granularity = gqa_ratio qhead_granularity = gqa_ratio
qlen_granularity = tile_q // qhead_granularity qlen_granularity = tile_q // qhead_granularity
kvlen_granularity = max(128, self.page_size) kvlen_granularity = 128
block_size = self.page_size block_size = 1
qo_indptr_cpu = qo_indptr.to("cpu", dtype=torch.int32) qo_indptr_cpu = qo_indptr.to("cpu", dtype=torch.int32)
kv_indptr_cpu = kv_indptr.to("cpu", dtype=torch.int32) kv_indptr_cpu = kv_indptr.to("cpu", dtype=torch.int32)