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

Skip oow slot freeing under eagle (#34823)

Browse Source
This commit is contained in:
Ke Bao
2026-08-14 17:13:21 +08:00
committed by GitHub
parent 4d94f1d310
commit 18107e38d2
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/unified_radix_cache.py
+6 -1
View File
@@ -773,7 +773,12 @@ class UnifiedRadixCache(BasePrefixCache):
if cl is not None: if cl is not None:
effective_cache_len = min(effective_cache_len, cl) effective_cache_len = min(effective_cache_len, cl)
if envs.SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS.get(): # swa_evicted_seqlen is a raw-token length, but under EAGLE the insert key is
# bigram-indexed, so SWA would carve tombstones at the wrong offset (#34653).
if (
envs.SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS.get()
and not self.tree_core.is_eagle
):
for comp in self._components_tuple: for comp in self._components_tuple:
comp.free_out_of_window_slots( comp.free_out_of_window_slots(
req, effective_cache_len - 1, insert_params req, effective_cache_len - 1, insert_params