[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>
This commit is contained in:
Nick Hill
2026-07-09 18:17:11 +00:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 2285cfca46
commit ff8d3488f2
@@ -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