chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Revert "[HiCache] Enable spec v2 + decode KV cache offloading compatibility" (#19613)

Browse Source
This commit is contained in:
Kangyan-Zhou
2026-02-28 21:54:32 -08:00
committed by GitHub
parent 8167346609
commit dcf462cfba
3 changed files with 7 additions and 191 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py
View File

@@ -20,7 +20,6 @@ from sglang.srt.mem_cache.memory_pool_host import (
MLATokenToKVPoolHost,
)
from sglang.srt.server_args import ServerArgs
from sglang.srt.utils.common import ceil_align
if TYPE_CHECKING:
from sglang.srt.managers.schedule_batch import Req
@@ -192,18 +191,6 @@ class DecodeKVCacheOffloadManager:
# Free the incremental part of the request
self.token_to_kv_pool_allocator.free(kv_indices)
# Free over-allocated KV cache slots (e.g. from speculative decoding v2).
# Without spec v2, start_p == end_p so this is a no-op.
start_p, end_p = req.pop_overallocated_kv_cache()
if self.page_size > 1:
start_p = ceil_align(start_p, self.page_size)
if start_p < end_p:
overalloc_indices = self.req_to_token_pool.req_to_token[
req.req_pool_idx, start_p:end_p
]
self.token_to_kv_pool_allocator.free(overalloc_indices)
self.req_to_token_pool.free(req)
self.tree_cache.protected_size_ -= len(req.prefix_indices)