diff --git a/python/sglang/srt/managers/scheduler_pp_mixin.py b/python/sglang/srt/managers/scheduler_pp_mixin.py index 479b2322f..0a55601d4 100644 --- a/python/sglang/srt/managers/scheduler_pp_mixin.py +++ b/python/sglang/srt/managers/scheduler_pp_mixin.py @@ -714,12 +714,14 @@ class SchedulerPPMixin: seq_lens.append(len(input_ids)) latencies.append(latency_ms) - # Release KV cache + # Release KV and Mamba cache if req.req_pool_idx is not None: kv_indices = self.req_to_token_pool.req_to_token[ req.req_pool_idx, : req.extend_range.end ] 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) req.kv = None