From ff8d3488f248acc8b5c1d23243723eeb00c74914 Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Thu, 9 Jul 2026 19:17:11 +0100 Subject: [PATCH] [Bugfix][MRV2] Reset num_accepted_tokens on add_request in all modes (#48132) Signed-off-by: Nick Hill Co-authored-by: Claude Opus 4.8 (1M context) --- vllm/v1/worker/gpu/model_states/mamba_hybrid.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/vllm/v1/worker/gpu/model_states/mamba_hybrid.py b/vllm/v1/worker/gpu/model_states/mamba_hybrid.py index 26439a9858..d04400afae 100644 --- a/vllm/v1/worker/gpu/model_states/mamba_hybrid.py +++ b/vllm/v1/worker/gpu/model_states/mamba_hybrid.py @@ -100,12 +100,13 @@ class MambaHybridModelState(DefaultModelState): def add_request(self, req_index: int, new_req_data: NewRequestData) -> None: super().add_request(req_index, new_req_data) + # Must reset the speculative acceptance count in this idx which could be stale. + self.num_accepted_tokens_gpu[req_index] = 1 if self._align_mode: # Seed the running state block from the resumed/prefilled position. self._mamba_state_idx_gpu[req_index] = ( new_req_data.num_computed_tokens - 1 ) // self.cache_config.block_size - self.num_accepted_tokens_gpu[req_index] = 1 def _get_mamba_group_info( self, kv_cache_config: KVCacheConfig