Retain CUDA IPC events in MP adapter

[he-yufeng]

What this PR does / why we need it:

Fixes #3236.

LMCacheMPWorkerAdapter now keeps the producer-side CUDA event objects alive for pending MP store and retrieve requests. The adapter already tracks the matching futures; this adds matching event references and drops them when the future completes or when pending work is drained after the server becomes unhealthy.

Without retaining the event object, the daemon can receive an IPC handle whose producer-side event has already been collected, which can make torch.cuda.Event.from_ipc_handle(...) fail with CUDA error: invalid argument.

Special notes for your reviewers:

The unit tests use a fake CUDA event plus weakref to check both sides of the lifetime: pending requests keep the event alive, and get_finished() releases it once the matching future completes.

Validation run locally on Windows with a CPU/source-only test environment:

python -m ruff check lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
python -m ruff format --check lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
.\.venv\Scripts\python.exe -m pytest tests\v1\test_vllm_mp_adapter.py -q
.\.venv\Scripts\python.exe -m py_compile lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
git diff --check
git fsck --no-progress

If applicable:

• this PR contains user facing changes - docs added
• this PR contains unit tests

[gemini-code-assist]

Code Review

This pull request modifies the LMCacheMPWorkerAdapter to maintain references to CUDA event objects during store and retrieve operations, ensuring they are not garbage collected while their IPC handles are in use. It also includes regression tests to verify that these events are correctly held and released upon request completion. I have no feedback to provide.

[KuntaiDu]

LGTM! @ApostaC @YaoJiayi can you also check?

[he-yufeng]

Follow-up pushed in 6dd0685.

Changes:

• replaced direct torch.cuda.Event annotations with a small structural ipc_handle() protocol, so this adapter API no longer couples the type surface to torch.cuda;
• annotated the fake heartbeat callback slot to clear the mypy assignment error.

Validation on Windows:

• python -m py_compile lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
• git diff --check
• python -m ruff check lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
• python -m mypy --config-file pyproject.toml lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py

Note: python -m pytest tests\v1\test_vllm_mp_adapter.py -q is blocked locally before collection by the missing lmcache.c_ops native extension in this Windows checkout.

[he-yufeng]

DCO is fixed now. I rebased the two PR commits with Signed-off-by footers and force-pushed the same code diff; current head is 530fd55. The temporary bad squash was immediately reverted, and the PR diff is back to the intended two files.

[he-yufeng]

Rebased onto the latest dev and resolved the test conflict; current head is 383d485. The code diff is still limited to the MP adapter and its regression tests. Validation on Windows: python -m py_compile lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py; python -m ruff check lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py; python -m mypy --config-file pyproject.toml lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py; git diff --check upstream/dev...HEAD. Targeted pytest is still blocked before collection by missing lmcache.c_ops in this Windows checkout.

[he-yufeng]

Updated again in 3cd6b98.

Changes:

• rebased the PR onto current dev;
• fixed the CI failure in the new event-retention tests by stubbing the current TransferContext path instead of the older send_lmcache_request/to_cuda_future path;
• the earlier torch_dev review concern is addressed in 872da20: the adapter retains IPC-capable events through a small _IpcEvent protocol and no longer exposes a torch.cuda.Event type dependency in this adapter API.

Validation on Windows:

• python -m py_compile lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
• python -m ruff check lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
• python -m mypy --config-file pyproject.toml lmcache\integration\vllm\vllm_multi_process_adapter.py tests\v1\test_vllm_mp_adapter.py
• git diff --check

The targeted pytest command is still blocked locally before collection in this Windows checkout by the missing native lmcache.c_ops extension; the previous CI failure itself should be covered by the Linux test job after this push.

[he-yufeng]

Pushed 195edaa to fix the Linux CI failures in the new event-retention tests. The adapter was already clearing store_events / retrieve_events; the test mock's call_args was still holding the FakeCudaEvent reference, so the weakref assertion was testing the fixture rather than adapter cleanup. Validation on Windows: py_compile for the adapter and test, ruff check on touched files, mypy on touched files, and git diff --check passed. The targeted pytest still cannot collect locally on this Windows host because lmcache.c_ops is not built.

[he-yufeng]

Follow-up for the CI failure: the test was still retaining the fake CUDA event through the parent transfer_ctx mock's call history. I switched the cleanup to reset the parent mock, so the weakref check now verifies the adapter no longer owns the event after get_finished() removes it.

Validated locally:

• python -m py_compile tests\v1\test_vllm_mp_adapter.py
• ruff check tests\v1\test_vllm_mp_adapter.py
• mypy tests\v1\test_vllm_mp_adapter.py
• git diff --check

python -m pytest tests\v1\test_vllm_mp_adapter.py -q is still blocked in my Windows checkout before test collection by ModuleNotFoundError: No module named 'lmcache.c_ops' from tests/conftest.py.

[maobaolong]

LGTM.

[maobaolong]

@hlin99 Would you like to take a double check?

[hlin99]

LGTM