[LLM] Fix dependency graph paper cuts in serve llm

- Move CloudFileSystem import to module top level in cloud_downloader.py (was inline inside method)
- Move KVConnectorBackendFactory import to lazy local import inside _setup_kv_connector_backend() to avoid llm_config depending on engine-specific components at module load time

Signed-off-by: ps2181 <hellopritam31@gmail.com>

Author: ps2181

python/ray/llm/_internal/common/callbacks/cloud_downloader.py

@@ -4,6 +4,7 @@
 
 from pydantic import BaseModel, field_validator
 
+from ray.llm._internal.common.utils.cloud_utils import CloudFileSystem
 from .base import CallbackBase
 
 logger = logging.getLogger(__name__)
@@ -74,8 +75,6 @@ def __init__(self, **kwargs: Any) -> None:
 
     def on_before_download_model_files_distributed(self) -> None:
         """Download files from cloud storage to local paths before model files are downloaded."""
-        from ray.llm._internal.common.utils.cloud_utils import CloudFileSystem
-
         paths = self.kwargs["paths"]
         start_time = time.monotonic()
         for cloud_uri, local_path in paths:

python/ray/llm/_internal/serve/core/configs/llm_config.py

@@ -44,9 +44,6 @@
     TPUConfig,
     infer_hardware_kind_from_bundles,
 )
-from ray.llm._internal.serve.engines.vllm.kv_transfer.factory import (
-    KVConnectorBackendFactory,
-)
 from ray.llm._internal.serve.observability.logging import get_logger
 from ray.serve._private.config import DeploymentConfig, handle_num_replicas_auto
 
@@ -609,6 +606,10 @@ def _setup_kv_connector_backend(self):
             raise ValueError("Connector type is not specified.")
 
         # 2. Setup the backend using factory
+        from ray.llm._internal.serve.engines.vllm.kv_transfer.factory import (
+            KVConnectorBackendFactory,
+        )
+
         kv_connector_backend = KVConnectorBackendFactory.create_backend(
             kv_connector, self
         )