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

[bug6] clear stale mamba cow source on rematch (#29354)

Browse Source 
Co-authored-by: Zhangheng <hzh0425@apache.org>
Co-authored-by: Hanming Lu <hanminglu@meta.com>
Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com>
This commit is contained in:
Feng Yao
2026-07-02 13:04:09 +08:00
committed by GitHub
co-authored by Zhangheng Hanming Lu Hanming Lu
parent 307094dc7d
commit d15b79e96e
1 changed files with 12 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler.py
+12 -9
View File
@@ -2895,15 +2895,18 @@ class Scheduler(
# pre-existing from a session). Session-held slots have their own
# lifecycle and freeing them here causes double-free.
added = len(adder.can_run_list) > 0 and req is adder.can_run_list[-1]
if (
not added
and req.mamba_pool_idx is not None
and not getattr(req, "session", None)
):
self.tree_cache.req_to_token_pool.mamba_allocator.free(
req.mamba_pool_idx.unsqueeze(-1)
)
req.mamba_pool_idx = None
if not added:
# init_next_round_input() may stage deferred Mamba COW/clear
# metadata before add_one_req() rejects the request.
req.mamba_cow_src_index = None
req.mamba_needs_clear = False
if req.mamba_pool_idx is not None and not getattr(
req, "session", None
):
self.tree_cache.req_to_token_pool.mamba_allocator.free(
req.mamba_pool_idx.unsqueeze(-1)
)
req.mamba_pool_idx = None
break
if mamba_allocator is not None: