Tighten non-GPU multiprocess adapter typing (replace Any with concrete IPC/MQ types)

[Copilot]

This updates the non-GPU multiprocess adapter interfaces to use concrete protocol types instead of Any where appropriate, and aligns affected tests with the stricter key contract. The goal is to make the IPC boundary explicitly typed across base and concrete adapters without introducing a dependency on lmc_ops GPU format typing.

• Adapter base typing hardening (adapter_connector/base.py)

  • mq_client now uses MessageQueueClient.
  • Abstract method key params now use IPCCacheEngineKey (prepare_store, commit_store, prepare_retrieve, commit_retrieve).
  • layout_hints now uses LayoutHints | None.
  • gpu_kv_format remains Any | None (no lmc_ops.GPUKVFormat typing import).
  • compute_kv_layout GPU format return element remains Any.

• Concrete adapter signature alignment

  • adapter_connector/shm.py: updated mq_client and all key signatures to MessageQueueClient / IPCCacheEngineKey.
  • adapter_connector/pickle.py: same signature updates as SHM path.

• Test key contract update (test_non_cuda_data_transfer.py)

  • Replaced plain string key usage in test_non_gpu_context_shm_store_retrieve_flow_with_mocked_mq with a proper IPCCacheEngineKey instance (_default_key()), and reused it across prepare/commit store/retrieve calls.

Example of the key change pattern:

key = _default_key()
store_result = context.prepare_store(key=key, instance_id=1)
assert context.commit_store(key, 1, store_views)
retrieve_views = context.prepare_retrieve(key=key, instance_id=1)
assert context.commit_retrieve(key, 1)

Original prompt

Replace all Any type annotations in lmcache/v1/multiprocess/adapter_connector/base.py with concrete types:

1. mq_client: Any → mq_client: MessageQueueClient (import from lmcache.v1.multiprocess.mq)
2. key: Any in abstract methods (prepare_store, commit_store, prepare_retrieve, commit_retrieve) → key: IPCCacheEngineKey (import from lmcache.v1.multiprocess.custom_types)
3. layout_hints: Any | None → layout_hints: LayoutHints | None (import from lmcache.v1.gpu_connector.utils)
4. gpu_kv_format: Any | None → gpu_kv_format: "lmc_ops.GPUKVFormat" | None (use TYPE_CHECKING import for lmcache.c_ops)
5. Return type of compute_kv_layout: tuple[int, int, int, str, Any] → tuple[int, int, int, str, "lmc_ops.GPUKVFormat"]

Also update lmcache/v1/multiprocess/adapter_connector/shm.py which has key: Any in its method signatures — change to key: IPCCacheEngineKey.

Also check lmcache/v1/multiprocess/adapter_connector/pickle.py (if it exists at that path or similar) for the same key: Any and mq_client: Any patterns.

For the test file tests/v1/multiprocess/test_non_cuda_data_transfer.py, find the test test_non_gpu_context_shm_store_retrieve_flow_with_mocked_mq which passes key="k" (a plain string) to context.prepare_store(key="k", instance_id=1). Fix it to use a proper IPCCacheEngineKey instance instead of a string. Look at how IPCCacheEngineKey is constructed elsewhere in the test suite (e.g. test_free_locks.py uses create_cache_key(0)) and use a similar approach. The IPCCacheEngineKey likely needs fields like model_name, worker_id, request_id, and token hash info — check its definition and construct a minimal valid instance.

Similarly fix any other places in the same test file where a plain string is passed as a key to these methods.

The following is the prior conversation context from the user's chat exploration (may be truncated):

This pull request was created from Copilot chat.

[hlin99]

remove lmc_ops.GPUKVFormat changes, we can use Any for lmc_ops.GPUKVFormat, i don't want to import lmc_ops

[Copilot]

remove lmc_ops.GPUKVFormat changes, we can use Any for lmc_ops.GPUKVFormat, i don't want to import lmc_ops

Addressed in commit 9908844. I removed the lmc_ops.GPUKVFormat typing/import pattern in adapter_connector/base.py and switched those annotations back to Any.