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

[mem_cache] Build empty-prefix last_loc sentinel on-device to avoid per-call H2D sync (#32575)

Browse Source 
Signed-off-by: xingyuliu <charlotteliu12x@gmail.com>
This commit is contained in:
Xingyu Liu
2026-08-03 18:55:03 -07:00
committed by GitHub
parent a84e70eb1e
commit 572924634b
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/allocation.py
+1 -1
View File
@@ -352,7 +352,7 @@ def alloc_for_extend(
else: else:
# Paged allocation - build last_loc # Paged allocation - build last_loc
last_loc = [ last_loc = [
(t[-1:] if len(t) > 0 else torch.tensor([-1], device=batch.device)) (t[-1:] if len(t) > 0 else torch.full((1,), -1, device=batch.device))
for t in prefix_tensors for t in prefix_tensors
] ]
out_cache_loc = alloc_paged_token_slots_extend( out_cache_loc = alloc_paged_token_slots_extend(