Get the internal state of the object with the given key. Args: key: The object key. Returns: The L1ObjectSta

Swap the registry content in place with ``new_nodes``.

Check if key exists in storage.

Increase ref count for the given MemoryObj by one.

Wait for an event fd to be signaled. Returns True if signaled within timeout, False otherwise.

Get the shapes of the KV cache in LMCache

--- helpers ---

Wait for an event fd to be signaled.

Create a test ObjectKey with the given chunk ID.

skip stamps the given skip reason on the pod (without injecting), marshals it, and returns an Allowed patch response. The pod is still admitted (fail-

Wait for the future to be done. Args: timeout (Optional[float]): Maximum time to wait in seconds. If Non

=========================== Test helpers ===========================

Create a test ObjectKey with the given chunk ID.

Return a status dict describing L1 cache state.

Block the host until this event completes. No-op in the stub implementation.

Get or create the singleton instance. Args: config: Required for first-time initialization. Optional for subseque

Create a CacheEngineKey for testing.

Create a test TensorMemoryObj with the given size.

Create a TensorMemoryObj that references page(s) in the registered buffer. Args: buffer: The flat uint8 buffer registered with NIXL.

Wait for an event fd to be signaled. Returns True if signaled within timeout, False otherwise.

Blocking get for a single key.

Return per-group dtypes, or the legacy single dtype if the manager has not been registered yet (e.g. some unit tests).

Submit a put task to store KV cache to GDS asynchronously. :param on_complete_callback: Optional callback invoked after the GDS

Create a remote connector from the given URL. Supported URL formats: - redis://[[username]:[password]@]host[:port][/database][?option=va

Check if key exists.

Notify the eviction policy that new keys have been created. New keys are added as most recently used. Args: keys

Start the background prefetch loop thread.

Signal the loop to stop and wait for the thread to join. Cleans up any in-flight requests (releases L1 write locks, L2 locks

Write from a GPU buffer into the file. Args: src: Device pointer (``ctypes.c_void_p``) to write from. size: Number of

Allocate blocks of memory from the virtual address space of a given size and batch size. The actual allocated size could be larger th

Get the expected response class for a request type. Args: req_type: The request type to look up Returns: Expected respo

Read from the file into a GPU buffer. Args: dest: Device pointer (``ctypes.c_void_p``) to read into. size: Number of

Return a context manager that sets the active stream. Args: stream: The stream to activate. If None, uses the cur

Submit *fn* for execution on the worker determined by *affinity_key*. Returns a :class:`concurrent.futures.Future`.

Get ref count for the given MemoryObj.

Add a log line to the display. Thread-safe.

Submit a prefetch request for the given keys. Thread-safe. Can be called from any thread. The retained subset of found keys

Validate and set configuration value. Args: config: Configuration object to update. config_key: The configuration key to set.

Return ``{cache_salt: (sum, count)}`` for histogram *name*.

Create a MockL2Adapter with fast bandwidth.

Register a listener for L1Manager events. Args: listener: The listener to register.

Unsafe read the read-locked objects without adding new read locks. This method does not acquire read locks. Therefore, the caller need

Finish reading prefetched objects. Args: keys: Object keys that have been read. extra_count: Extra read locks to rele

Create a ZeroMQ socket with the specified protocol and role.

Reserve the object for writing into the storage manager. Args: keys (list[ObjectKey]): List of object keys to reserve fo

Start the background store loop thread.

Create a TensorMemoryObj that references page(s) in the registered buffer.

Yield concrete subclasses of *base_class* found by walking the submodules of *package*. Each subclass is yielded **at most once**, even when

Create a test ObjectKey with the given chunk ID.

Open the file and register the hipfile handle.

Submit a request to the server. Args: request_type (RequestType): The type of the request. request_payloads (list[Any

Allocate remote memory slots and register KV objects. Uses reservation-based admission: on the first batch for a request, reserves to

Create an L2 adapter using the factory registry. Looks up the type name for *config* via the config registry, then calls the matching factory

Load the config from a JSON file.

Check whether the device backend is available. Returns: False always, since this is a CPU-only stub.

Create an AdapterDescriptor for testing.

Create a small MemoryLayoutDesc for testing.

Return the registry-owned ``(host, port)`` matching the input. The returned tuple is taken from the registry itself, not from the cal

Store test data directly in L2 adapter and wait for completion.