diff --git a/tests/v1/kv_connector/unit/test_moriio_connector.py b/tests/v1/kv_connector/unit/test_moriio_connector.py index cb33bfbcf0..73cc5f0583 100644 --- a/tests/v1/kv_connector/unit/test_moriio_connector.py +++ b/tests/v1/kv_connector/unit/test_moriio_connector.py @@ -195,6 +195,17 @@ def _build_meta( return scheduler.build_connector_meta(scheduler_output) +def test_write_mode_does_not_reload_remote_kv_after_preemption(): + scheduler = _write_consumer_scheduler_for_finished_request() + request = create_request(request_id=1, num_tokens=32, do_remote_prefill=True) + + assert scheduler.get_num_new_matched_tokens(request, 0) == (32, True) + + assert request.kv_transfer_params is not None + request.kv_transfer_params["do_remote_prefill"] = False + assert scheduler.get_num_new_matched_tokens(request, 0) == (0, False) + + class FakeMoRIIOWrapper: # A fake MoRIIOWrapper for testing purposes def __init__(self, *args, **kwargs): diff --git a/tests/v1/kv_connector/unit/test_moriio_hma_scheduler.py b/tests/v1/kv_connector/unit/test_moriio_hma_scheduler.py index 1de729416e..4bd32f786a 100644 --- a/tests/v1/kv_connector/unit/test_moriio_hma_scheduler.py +++ b/tests/v1/kv_connector/unit/test_moriio_hma_scheduler.py @@ -824,7 +824,7 @@ def test_get_num_new_matched_tokens_write_plain_keeps_all_tokens(): req = SimpleNamespace( num_prompt_tokens=10, prompt_token_ids=list(range(10)), - kv_transfer_params=None, + kv_transfer_params={"do_remote_prefill": True}, ) n, is_async = sched.get_num_new_matched_tokens(req, num_computed_tokens=2) # Pure-attention WRITE: no N-1 drop; full length minus already-computed. @@ -848,7 +848,7 @@ def test_get_num_new_matched_tokens_supports_embeds_only_prompts( num_prompt_tokens=10, prompt_token_ids=None, prompt_embeds=object(), - kv_transfer_params=None, + kv_transfer_params={"do_remote_prefill": True}, ) assert sched.get_num_new_matched_tokens(req, num_computed_tokens) == ( diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/moriio/moriio_connector.py b/vllm/distributed/kv_transfer/kv_connector/v1/moriio/moriio_connector.py index 2d17d7a7ac..082643778e 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/moriio/moriio_connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/moriio/moriio_connector.py @@ -769,6 +769,9 @@ class MoRIIOConnectorScheduler: num_prompt_tokens = request.num_prompt_tokens num_external_tokens = max(num_prompt_tokens - num_computed_tokens, 0) if self.mode == MoRIIOMode.WRITE: + params = request.kv_transfer_params or {} + if not params.get("do_remote_prefill"): + return 0, False # MoriiO in write mode, no remote prefill. Hybrid models never get # here: register_kv_caches refuses WRITE mode for them, so there is # no recurrent-state accounting to do.