[Bugfix][MRV2] Reset num_accepted_tokens on add_request in all modes (#48132)
Signed-off-by: Nick Hill <nickhill123@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Nick Hill committed
Jul 9, 2026 at 19:17 UTC
ff8d3488f248acc8b5c1d23243723eeb00c74914
1 file changed
+2
-1
vllm/v1/worker/gpu/model_states/mamba_hybrid.py
+2
-1
@@ -100,12 +100,13 @@ class MambaHybridModelState(DefaultModelState):
100
101
def add_request(self, req_index: int, new_req_data: NewRequestData) -> None:
102
super().add_request(req_index, new_req_data)
103
+ # Must reset the speculative acceptance count in this idx which could be stale.
104
+ self.num_accepted_tokens_gpu[req_index] = 1
105
if self._align_mode:
106
# Seed the running state block from the resumed/prefilled position.
107
self._mamba_state_idx_gpu[req_index] = (
108
new_req_data.num_computed_tokens - 1
109
) // self.cache_config.block_size
108
- self.num_accepted_tokens_gpu[req_index] = 1
110
111
def _get_mamba_group_info(
112
self, kv_cache_config: KVCacheConfig