[Misc] Set default kv_buffer_device in a better way (#36862)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -13,6 +13,12 @@ KVConsumer = Literal["kv_consumer", "kv_both"]
|
||||
KVRole = Literal[KVProducer, KVConsumer]
|
||||
|
||||
|
||||
def kv_buffer_device_default_factory() -> str:
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
return current_platform.device_type
|
||||
|
||||
|
||||
@config
|
||||
class KVTransferConfig:
|
||||
"""Configuration for distributed KV cache transfer."""
|
||||
@@ -24,7 +30,7 @@ class KVTransferConfig:
|
||||
engine_id: str | None = None
|
||||
"""The engine id for KV transfers."""
|
||||
|
||||
kv_buffer_device: str | None = None
|
||||
kv_buffer_device: str = field(default_factory=kv_buffer_device_default_factory)
|
||||
"""The device used by kv connector to buffer the KV cache. Choices are
|
||||
'cuda', 'cpu' and 'xpu'."""
|
||||
|
||||
@@ -100,11 +106,6 @@ class KVTransferConfig:
|
||||
f"is set, supported roles are {get_args(KVRole)}"
|
||||
)
|
||||
|
||||
if self.kv_buffer_device is None:
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
self.kv_buffer_device = current_platform.device_type
|
||||
|
||||
@property
|
||||
def is_kv_transfer_instance(self) -> bool:
|
||||
return self.kv_connector is not None and self.kv_role in get_args(KVRole)
|
||||
|
||||
@@ -998,10 +998,7 @@ class NixlConnectorWorker:
|
||||
|
||||
# KV Caches and nixl tracking data.
|
||||
self.device_type = current_platform.device_type
|
||||
kv_buffer_device = vllm_config.kv_transfer_config.kv_buffer_device
|
||||
if kv_buffer_device is None:
|
||||
raise ValueError("kv_buffer_device must be set for NixlConnector")
|
||||
self.kv_buffer_device: str = kv_buffer_device
|
||||
self.kv_buffer_device: str = vllm_config.kv_transfer_config.kv_buffer_device
|
||||
if self.device_type not in _NIXL_SUPPORTED_DEVICE:
|
||||
raise RuntimeError(f"{self.device_type} is not supported.")
|
||||
elif self.kv_buffer_device not in _NIXL_SUPPORTED_DEVICE[self.device_type]:
|
||||
|
||||
Reference in New Issue
Block a user