minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[Bugfix] Release Mamba cache after PP dynamic chunk profiling (#31321)

Browse Source
This commit is contained in:
Xuwei
2026-07-16 02:32:07 +08:00
committed by GitHub
parent 2d00e20a52
commit 7e7129acd7
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler_pp_mixin.py
+3 -1
View File
@@ -714,12 +714,14 @@ class SchedulerPPMixin:
seq_lens.append(len(input_ids)) seq_lens.append(len(input_ids))
latencies.append(latency_ms) latencies.append(latency_ms)
# Release KV cache # Release KV and Mamba cache
if req.req_pool_idx is not None: if req.req_pool_idx is not None:
kv_indices = self.req_to_token_pool.req_to_token[ kv_indices = self.req_to_token_pool.req_to_token[
req.req_pool_idx, : req.extend_range.end req.req_pool_idx, : req.extend_range.end
] ]
self.token_to_kv_pool_allocator.free(kv_indices) self.token_to_kv_pool_allocator.free(kv_indices)
if req.mamba_pool_idx is not None:
self.req_to_token_pool.free_mamba_cache(req)
self.req_to_token_pool.free(req) self.req_to_token_pool.free(req)
req.kv = None req.kv = None