vllm.distributed.kv_transfer
Modules:
Name | Description |
---|---|
kv_connector |
|
kv_connector_agent |
A centralized entrypoint to perform distributed KV cache transfer. |
kv_lookup_buffer |
|
kv_pipe |
|
kv_transfer_state |
|
KVConnectorBaseType
module-attribute
¶
KVConnectorBaseType = Union[
KVConnectorBase, KVConnectorBase_V1
]
__all__
module-attribute
¶
__all__ = [
"get_kv_transfer_group",
"has_kv_transfer_group",
"is_v1_kv_transfer_group",
"ensure_kv_transfer_initialized",
"KVConnectorBaseType",
]
ensure_kv_transfer_initialized
¶
ensure_kv_transfer_initialized(
vllm_config: VllmConfig,
) -> None
Initialize KV cache transfer parallel group.
Source code in vllm/distributed/kv_transfer/kv_transfer_state.py
get_kv_transfer_group
¶
get_kv_transfer_group() -> KVConnectorBaseType
is_v1_kv_transfer_group
¶
is_v1_kv_transfer_group(
connector: Optional[KVConnectorBaseType] = None,
) -> bool
Check if the KV connector is the v1 connector. If the argument is None, it will check the global KV connector
Parameters:
Name | Type | Description | Default |
---|---|---|---|
connector
|
Optional[KVConnectorBaseType]
|
The KV connector to check. If None, it will check the global KV connector. |
None
|
Note
This function will no-longer be needed after the v1 KV connector becomes the default.