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

[AMD] fix aiter cannot get heuristic kernel regression (#37438)

Browse Source
This commit is contained in:
billishyahao
2026-09-01 08:35:31 -07:00
committed by GitHub
parent cb6dd58fbe
commit 44a92e54b9
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/aiter_backend.py
+10 -3
View File
@@ -466,9 +466,16 @@ class AiterAttnBackend(AttentionBackend):
intra_batch_mode = False intra_batch_mode = False
# Zero-pad topology (h12->qh16): prefer Gluon decode over PS kernel. # Zero-pad topology (h12->qh16): prefer Gluon decode over PS kernel.
if self.head_pad_mode == "zero" and self.kv_cache_dtype == fp8_dtype: if self.head_pad_mode == "zero" and self.kv_cache_dtype == fp8_dtype:
_use_mla_ps_kernel = False # Disable ps only when gluon kernel is selected to avoid falling
fast_mode = False # back to incorrect aiter kernel
intra_batch_mode = False if prefer_mla_gluon_decode(
head_pad_mode=self.head_pad_mode,
num_head=self.num_head,
kv_cache_dtype=self.kv_cache_dtype,
):
_use_mla_ps_kernel = False
fast_mode = False
intra_batch_mode = False
log_mla_gluon_capability(logger) log_mla_gluon_capability(logger)
self.max_split_per_batch = 32 if _use_mla_ps_kernel else None self.max_split_per_batch = 32 if _use_mla_ps_kernel else None