A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 21:48 UTC
[Cleanup] Remove unused `KVConnectorModelRunnerMixin` methods (#32077)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Jan 20, 2026 at 19:16 UTC
6f067b1fb7309ed9dc05f187f25ca29f82f6fd9a
2 files changed
+2
-32
tests/v1/kv_connector/unit/test_nixl_connector.py
+2
-2
@@ -518,7 +518,7 @@ class TestNixlHandshake:
518
)
519
connector.bind_connector_metadata(metadata)
520
521
- # Mimic maybe_setup_kv_connector in gpu_model_runner.
521
+ # Mimic logic in KVConnectorModelRunnerMixin._get_kv_connector_output.
522
dummy_ctx = ForwardContext(
523
no_compile_layers={},
524
attn_metadata={},
@@ -531,7 +531,7 @@ class TestNixlHandshake:
531
f"start_load_kv took {_after_load - _before_load} seconds"
532
)
533
534
- # Mimic get_finished_kv_transfers in gpu_model_runner.
534
+ # Mimic logic in KVConnectorModelRunnerMixin._get_kv_connector_output.
535
_, done_recving = connector.get_finished(finished_req_ids=set())
536
if len(done_recving) > 0:
537
assert request_id in done_recving
vllm/v1/worker/kv_connector_model_runner_mixin.py
-30
@@ -38,42 +38,12 @@ logger = init_logger(__name__)
38
39
# Defined as a kv connector functionality mixin for ModelRunner (GPU, TPU)
40
class KVConnectorModelRunnerMixin:
41
- @staticmethod
42
- def maybe_setup_kv_connector(scheduler_output: "SchedulerOutput"):
43
- # Update KVConnector with the KVConnector metadata forward().
44
- if has_kv_transfer_group():
45
- kv_connector = get_kv_transfer_group()
46
- assert isinstance(kv_connector, KVConnectorBase)
47
- assert scheduler_output.kv_connector_metadata is not None
48
- kv_connector.bind_connector_metadata(scheduler_output.kv_connector_metadata)
49
-
50
- # Background KV cache transfers happen here.
51
- # These transfers are designed to be async and the requests
52
- # involved may be disjoint from the running requests.
53
- # Do this here to save a collective_rpc.
54
- kv_connector.start_load_kv(get_forward_context())
55
-
41
@staticmethod
42
def ensure_kv_transfer_shutdown() -> None:
43
# has_kv_transfer_group can be None during interpreter shutdown.
44
if has_kv_transfer_group and has_kv_transfer_group(): # type: ignore[truthy-function]
45
ensure_kv_transfer_shutdown()
46
62
- @staticmethod
63
- def maybe_wait_for_kv_save() -> None:
64
- if has_kv_transfer_group():
65
- get_kv_transfer_group().wait_for_save()
66
-
67
- @staticmethod
68
- def get_finished_kv_transfers(
69
- scheduler_output: "SchedulerOutput",
70
- ) -> tuple[set[str] | None, set[str] | None]:
71
- if has_kv_transfer_group():
72
- return get_kv_transfer_group().get_finished(
73
- scheduler_output.finished_req_ids
74
- )
75
- return None, None
76
-
47
@staticmethod
48
def kv_connector_no_forward(
49
scheduler_output: "SchedulerOutput", vllm_config: VllmConfig