From e49e9711317618218289e479660e28cbef555514 Mon Sep 17 00:00:00 2001 From: Hongsheng Jin Date: Tue, 6 Oct 2026 16:40:35 +0800 Subject: [PATCH] [mem_cache] Keep prefix_indices current for chunked requests kept out of the tree (#42467) Co-authored-by: Claude Opus 5.5 Co-authored-by: hnyls2002 Co-authored-by: Liangsheng Yin --- python/sglang/srt/mem_cache/common.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/python/sglang/srt/mem_cache/common.py b/python/sglang/srt/mem_cache/common.py index db6241b6d2..6131f6f32b 100644 --- a/python/sglang/srt/mem_cache/common.py +++ b/python/sglang/srt/mem_cache/common.py @@ -176,6 +176,10 @@ def checkpoint_kv_cache(req: Req, tree_cache: BasePrefixCache) -> None: # insert; a finished request belongs in release_kv_cache. assert not req.finished(), f"checkpointing finished request {req.rid}" if req.skip_radix_cache_insert: + # Kept out of the tree; the next extend still resumes from prefix_indices. + req.prefix_indices = tree_cache.req_to_token_pool.req_to_token[ + req.kv.req_pool_idx, : req.extend_range.end + ].to(dtype=torch.int64, copy=True) return tree_cache.checkpoint(req, up_to=req.extend_range.end)