[KV Offloading] Remove dummy worker-side stats from OffloadingConnector (#45905)

Signed-off-by: Alex <alex.tech.lab@outlook.com> Signed-off-by: AlexHuang <jihuihuang@alexai.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>

AlexHuang committed Jun 18, 2026 at 13:13 UTC 421c1ec4483b1db2cc6723a568518dc719ffe837

2 files changed -21

tests/v1/kv_connector/unit/test_offloading_connector.py

-16

@@ -383,22 +383,6 @@ def test_cpu_offloading_metrics() -> None:

383 total += sample.value

384 return total

385

386 - # Stats are drained asynchronously — if the transfer finishes

387 - # after the last engine step for that generate() call, the metrics

388 - # won't appear until a subsequent step. Retry with dummy generates

389 - # to force additional stats drains.

390 - deadline = time.monotonic() + _RESET_CACHE_TIMEOUT

391 - while time.monotonic() < deadline:

392 - store_bytes = _get_counter_value("vllm:kv_offload_store_bytes")

393 - load_bytes = _get_counter_value("vllm:kv_offload_load_bytes")

394 - if store_bytes > 0 and load_bytes > 0:

395 - break

396 - llm.generate(

397 - [TokensPrompt(prompt_token_ids=[0])],

398 - SamplingParams(max_tokens=1),

399 - use_tqdm=False,

400 - )

401 -

386 # New flat counter metrics

387 store_bytes = _get_counter_value("vllm:kv_offload_store_bytes")

388 assert store_bytes > 0, f"Expected store_bytes > 0, got {store_bytes}"

vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py

-5

@@ -190,11 +190,6 @@ class OffloadingConnector(KVConnectorBase_V1, SupportsHMA):

190 def get_kv_connector_stats(self) -> KVConnectorStats | None:

191 if self.connector_scheduler is not None:

192 return self.connector_scheduler.get_stats()

193 -

194 - # TODO(orozery): Remove once PR #43877 lands

195 - if self.connector_worker is not None:

196 - return OffloadingConnectorStats()

197 -

193 return None

194

195 @classmethod