diff --git a/vllm/config/kv_transfer.py b/vllm/config/kv_transfer.py index eb6116d0c0..172b7a8059 100644 --- a/vllm/config/kv_transfer.py +++ b/vllm/config/kv_transfer.py @@ -24,9 +24,9 @@ class KVTransferConfig: engine_id: str | None = None """The engine id for KV transfers.""" - kv_buffer_device: str = "cuda" - """The device used by kv connector to buffer the KV cache. Choices are - 'cuda' and 'cpu'.""" + kv_buffer_device: str | None = None + """The device used by kv connector to buffer the KV cache. Choices are + 'cuda','cpu' and 'xpu'.""" kv_buffer_size: float = 1e9 """The buffer size for TorchDistributedConnector. Measured in number of @@ -100,6 +100,11 @@ 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) diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py index 356a837fb3..f6ad03ba99 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py @@ -998,7 +998,10 @@ class NixlConnectorWorker: # KV Caches and nixl tracking data. self.device_type = current_platform.device_type - self.kv_buffer_device: str = vllm_config.kv_transfer_config.kv_buffer_device + 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 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]: