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

[bugfix] Add prev_prefix_len parameter to HiMambaRadixCache's _insert_helper() (#20539)

Browse Source
This commit is contained in:
ziruiliu
2026-03-14 09:54:14 +08:00
committed by GitHub
parent 4659b08fcf
commit dfd0a77a9a
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/hi_mamba_radix_cache.py
+4 -4
View File
@@ -647,6 +647,7 @@ class HiMambaRadixCache(MambaRadixCache):
value, value,
mamba_value, mamba_value,
chunked: bool = False, chunked: bool = False,
prev_prefix_len: int = 0,
) -> Tuple[int, bool]: ) -> Tuple[int, bool]:
assert mamba_value is not None, "Mamba value should not be None here." assert mamba_value is not None, "Mamba value should not be None here."
node.last_access_time = get_last_access_time() node.last_access_time = get_last_access_time()
@@ -677,12 +678,11 @@ class HiMambaRadixCache(MambaRadixCache):
node = new_node node = new_node
if node.evicted: if node.evicted:
# Unevicted nodes take ownership of the request's KV pages.
# Do NOT count them in total_prefix_length, otherwise
# cache_finished_req / cache_unfinished_req will free those
# pages even though the tree now references them.
self._unevict_node(node, value[:prefix_len]) self._unevict_node(node, value[:prefix_len])
else: else:
if prev_prefix_len < total_prefix_length + prefix_len:
start = max(0, prev_prefix_len - total_prefix_length)
self.token_to_kv_pool_allocator.free(value[start:prefix_len])
total_prefix_length += prefix_len total_prefix_length += prefix_len
self._inc_hit_count(node, chunked) self._inc_hit_count(node, chunked)