diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py b/vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py index 806267dcf3..6417bb472a 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/offloading/scheduler.py @@ -1146,6 +1146,14 @@ class OffloadingConnectorScheduler: def update_state_after_alloc( self, request: Request, blocks: KVCacheBlocks, num_external_tokens: int ): + # Track allocated block IDs regardless of num_external_tokens so + # re-allocation flush detection sees blocks allocated by other + # connectors in a MultiConnector setup. + for group_blocks in blocks.blocks: + self._current_batch_allocated_block_ids.update( + block.block_id for block in group_blocks if block.block_id != 0 + ) + if num_external_tokens == 0: return @@ -1168,10 +1176,6 @@ class OffloadingConnectorScheduler: req_status.group_states, ): group_blocks = blocks.blocks[group_config.group_idx] - self._current_batch_allocated_block_ids.update( - block.block_id for block in group_blocks if block.block_id != 0 - ) - tokens_per_block = group_config.tokens_per_block tokens_per_chunk = group_config.tokens_per_chunk offload_keys = group_state.offload_keys