chore: import upstream snapshot with attribution

python/ray/llm/_internal/common/utils/cloud_filesystem/__init__.py

@@ -0,0 +1,29 @@
+"""Cloud filesystem module for provider-specific implementations.
+
+This module provides a unified interface for cloud storage operations across
+different providers (S3, GCS, Azure) while allowing provider-specific optimizations.
+"""
+
+from ray.llm._internal.common.utils.cloud_filesystem.azure_filesystem import (
+    AzureFileSystem,
+)
+from ray.llm._internal.common.utils.cloud_filesystem.base import (
+    BaseCloudFileSystem,
+)
+from ray.llm._internal.common.utils.cloud_filesystem.gcs_filesystem import (
+    GCSFileSystem,
+)
+from ray.llm._internal.common.utils.cloud_filesystem.pyarrow_filesystem import (
+    PyArrowFileSystem,
+)
+from ray.llm._internal.common.utils.cloud_filesystem.s3_filesystem import (
+    S3FileSystem,
+)
+
+__all__ = [
+    "BaseCloudFileSystem",
+    "PyArrowFileSystem",
+    "GCSFileSystem",
+    "AzureFileSystem",
+    "S3FileSystem",
+]

python/ray/llm/_internal/common/utils/cloud_filesystem/azure_filesystem.py

@@ -0,0 +1,82 @@
+"""Azure-specific filesystem implementation.
+
+This module provides an Azure-specific implementation that delegates to PyArrowFileSystem.
+This maintains backward compatibility while allowing for future optimizations using
+native Azure tools (azcopy, azure-storage-blob SDK).
+"""
+
+from typing import List, Optional, Union
+
+from ray.llm._internal.common.utils.cloud_filesystem.base import BaseCloudFileSystem
+from ray.llm._internal.common.utils.cloud_filesystem.pyarrow_filesystem import (
+    PyArrowFileSystem,
+)
+
+
+class AzureFileSystem(BaseCloudFileSystem):
+    """Azure-specific implementation of cloud filesystem operations.
+
+    **Note**: This implementation currently delegates to PyArrowFileSystem to maintain
+    stability. Optimized implementation using azure-storage-blob SDK and azcopy
+    will be added in a future PR.
+    """
+
+    @staticmethod
+    def get_file(
+        object_uri: str, decode_as_utf_8: bool = True
+    ) -> Optional[Union[str, bytes]]:
+        """Download a file from cloud storage into memory.
+
+        Args:
+            object_uri: URI of the file (abfss:// or azure://)
+            decode_as_utf_8: If True, decode the file as UTF-8
+
+        Returns:
+            File contents as string or bytes, or None if file doesn't exist
+        """
+        return PyArrowFileSystem.get_file(object_uri, decode_as_utf_8)
+
+    @staticmethod
+    def list_subfolders(folder_uri: str) -> List[str]:
+        """List the immediate subfolders in a cloud directory.
+
+        Args:
+            folder_uri: URI of the directory (abfss:// or azure://)
+
+        Returns:
+            List of subfolder names (without trailing slashes)
+        """
+        return PyArrowFileSystem.list_subfolders(folder_uri)
+
+    @staticmethod
+    def download_files(
+        path: str,
+        bucket_uri: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+    ) -> None:
+        """Download files from cloud storage to a local directory.
+
+        Args:
+            path: Local directory where files will be downloaded
+            bucket_uri: URI of cloud directory
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude certain files from download (e.g .safetensors)
+        """
+        PyArrowFileSystem.download_files(
+            path, bucket_uri, substrings_to_include, suffixes_to_exclude
+        )
+
+    @staticmethod
+    def upload_files(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload files to cloud storage.
+
+        Args:
+            local_path: The local path of the files to upload.
+            bucket_uri: The bucket uri to upload the files to, must start with
+                `abfss://` or `azure://`.
+        """
+        PyArrowFileSystem.upload_files(local_path, bucket_uri)

python/ray/llm/_internal/common/utils/cloud_filesystem/base.py

@@ -0,0 +1,81 @@
+"""Abstract base class for cloud filesystem implementations.
+
+This module defines the interface that all cloud storage provider implementations
+must follow, ensuring consistency across different providers while allowing
+provider-specific optimizations.
+"""
+
+from abc import ABC, abstractmethod
+from typing import List, Optional, Union
+
+
+class BaseCloudFileSystem(ABC):
+    """Abstract base class for cloud filesystem implementations.
+
+    This class defines the interface that all cloud storage provider implementations
+    must implement. Provider-specific classes (S3FileSystem, GCSFileSystem, etc.)
+    will inherit from this base class and provide optimized implementations for
+    their respective cloud storage platforms.
+    """
+
+    @staticmethod
+    @abstractmethod
+    def get_file(
+        object_uri: str, decode_as_utf_8: bool = True
+    ) -> Optional[Union[str, bytes]]:
+        """Download a file from cloud storage into memory.
+
+        Args:
+            object_uri: URI of the file (s3://, gs://, abfss://, or azure://)
+            decode_as_utf_8: If True, decode the file as UTF-8
+
+        Returns:
+            File contents as string or bytes, or None if file doesn't exist
+        """
+        pass
+
+    @staticmethod
+    @abstractmethod
+    def list_subfolders(folder_uri: str) -> List[str]:
+        """List the immediate subfolders in a cloud directory.
+
+        Args:
+            folder_uri: URI of the directory (s3://, gs://, abfss://, or azure://)
+
+        Returns:
+            List of subfolder names (without trailing slashes)
+        """
+        pass
+
+    @staticmethod
+    @abstractmethod
+    def download_files(
+        path: str,
+        bucket_uri: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+    ) -> None:
+        """Download files from cloud storage to a local directory.
+
+        Args:
+            path: Local directory where files will be downloaded
+            bucket_uri: URI of cloud directory
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude certain files from download (e.g .safetensors)
+        """
+        pass
+
+    @staticmethod
+    @abstractmethod
+    def upload_files(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload files to cloud storage.
+
+        Args:
+            local_path: The local path of the files to upload.
+            bucket_uri: The bucket uri to upload the files to, must start with
+                `s3://`, `gs://`, `abfss://`, or `azure://`.
+        """
+        pass

python/ray/llm/_internal/common/utils/cloud_filesystem/gcs_filesystem.py

@@ -0,0 +1,81 @@
+"""GCS-specific filesystem implementation.
+
+This module provides a GCS-specific implementation.
+This maintains backward compatibility while allowing for future optimizations using
+native GCS tools (gsutil, google-cloud-storage SDK).
+"""
+
+from typing import List, Optional, Union
+
+from ray.llm._internal.common.utils.cloud_filesystem.base import BaseCloudFileSystem
+from ray.llm._internal.common.utils.cloud_filesystem.pyarrow_filesystem import (
+    PyArrowFileSystem,
+)
+
+
+class GCSFileSystem(BaseCloudFileSystem):
+    """GCS-specific implementation of cloud filesystem operations.
+
+    **Note**: This implementation currently delegates to PyArrowFileSystem to maintain
+    stability. Optimized implementation using google-cloud-storage SDK and gsutil
+    will be added in a future PR.
+    """
+
+    @staticmethod
+    def get_file(
+        object_uri: str, decode_as_utf_8: bool = True
+    ) -> Optional[Union[str, bytes]]:
+        """Download a file from cloud storage into memory.
+
+        Args:
+            object_uri: URI of the file (gs://)
+            decode_as_utf_8: If True, decode the file as UTF-8
+
+        Returns:
+            File contents as string or bytes, or None if file doesn't exist
+        """
+        return PyArrowFileSystem.get_file(object_uri, decode_as_utf_8)
+
+    @staticmethod
+    def list_subfolders(folder_uri: str) -> List[str]:
+        """List the immediate subfolders in a cloud directory.
+
+        Args:
+            folder_uri: URI of the directory (gs://)
+
+        Returns:
+            List of subfolder names (without trailing slashes)
+        """
+        return PyArrowFileSystem.list_subfolders(folder_uri)
+
+    @staticmethod
+    def download_files(
+        path: str,
+        bucket_uri: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+    ) -> None:
+        """Download files from cloud storage to a local directory.
+
+        Args:
+            path: Local directory where files will be downloaded
+            bucket_uri: URI of cloud directory
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude certain files from download (e.g .safetensors)
+        """
+        PyArrowFileSystem.download_files(
+            path, bucket_uri, substrings_to_include, suffixes_to_exclude
+        )
+
+    @staticmethod
+    def upload_files(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload files to cloud storage.
+
+        Args:
+            local_path: The local path of the files to upload.
+            bucket_uri: The bucket uri to upload the files to, must start with `gs://`.
+        """
+        PyArrowFileSystem.upload_files(local_path, bucket_uri)

python/ray/llm/_internal/common/utils/cloud_filesystem/pyarrow_filesystem.py

@@ -0,0 +1,389 @@
+"""PyArrow-based filesystem implementation for cloud storage.
+
+This module provides a PyArrow-based implementation of the cloud filesystem
+interface, supporting S3, GCS, and Azure storage providers.
+"""
+
+import os
+from concurrent.futures import ThreadPoolExecutor
+from typing import List, Optional, Tuple, Union
+from urllib.parse import urlparse
+
+import pyarrow.fs as pa_fs
+
+from ray.llm._internal.common.observability.logging import get_logger
+from ray.llm._internal.common.utils.cloud_filesystem.base import BaseCloudFileSystem
+
+logger = get_logger(__name__)
+
+
+class PyArrowFileSystem(BaseCloudFileSystem):
+    """PyArrow-based implementation of cloud filesystem operations.
+
+    This class provides a unified interface for cloud storage operations using
+    PyArrow's filesystem abstraction. It supports S3, GCS, and Azure storage
+    providers.
+    """
+
+    @staticmethod
+    def get_fs_and_path(object_uri: str) -> Tuple[pa_fs.FileSystem, str]:
+        """Get the appropriate filesystem and path from a URI.
+
+        Args:
+            object_uri: URI of the file (s3://, gs://, abfss://, or azure://)
+                If URI contains 'anonymous@', anonymous access is used.
+                Example: s3://anonymous@bucket/path
+
+        Returns:
+            Tuple of (filesystem, path)
+        """
+
+        if object_uri.startswith("pyarrow-"):
+            object_uri = object_uri[8:]
+
+        anonymous = False
+        # Check for anonymous access pattern (only for S3/GCS)
+        # e.g. s3://anonymous@bucket/path
+        if "@" in object_uri and not (
+            object_uri.startswith("abfss://") or object_uri.startswith("azure://")
+        ):
+            parts = object_uri.split("@", 1)
+            # Check if the first part ends with "anonymous"
+            if parts[0].endswith("anonymous"):
+                anonymous = True
+                # Remove the anonymous@ part, keeping the scheme
+                scheme = parts[0].split("://")[0]
+                object_uri = f"{scheme}://{parts[1]}"
+
+        if object_uri.startswith("s3://"):
+            endpoint = os.getenv("AWS_ENDPOINT_URL_S3", None)
+            virtual_hosted_style = os.getenv("AWS_S3_ADDRESSING_STYLE", None)
+            fs = pa_fs.S3FileSystem(
+                anonymous=anonymous,
+                endpoint_override=endpoint,
+                force_virtual_addressing=(virtual_hosted_style == "virtual"),
+            )
+            path = object_uri[5:]  # Remove "s3://"
+        elif object_uri.startswith("gs://"):
+            fs = pa_fs.GcsFileSystem(anonymous=anonymous)
+            path = object_uri[5:]  # Remove "gs://"
+        elif object_uri.startswith("abfss://"):
+            fs, path = PyArrowFileSystem._create_abfss_filesystem(object_uri)
+        elif object_uri.startswith("azure://"):
+            fs, path = PyArrowFileSystem._create_azure_filesystem(object_uri)
+        else:
+            raise ValueError(f"Unsupported URI scheme: {object_uri}")
+
+        return fs, path
+
+    @staticmethod
+    def _create_azure_filesystem(object_uri: str) -> Tuple[pa_fs.FileSystem, str]:
+        """Create an Azure filesystem for Azure Blob Storage or ABFSS.
+
+        Args:
+            object_uri: Azure URI (azure://container@account.blob.core.windows.net/path or
+                       abfss://container@account.dfs.core.windows.net/path)
+
+        Returns:
+            Tuple of (PyArrow FileSystem, path without scheme prefix)
+
+        Raises:
+            ImportError: If required dependencies are not installed.
+            ValueError: If the Azure URI format is invalid.
+        """
+        try:
+            import adlfs
+            from azure.identity import DefaultAzureCredential
+        except ImportError:
+            raise ImportError(
+                "You must `pip install adlfs azure-identity` "
+                "to use Azure/ABFSS URIs. "
+                "Note that these must be preinstalled on all nodes in the Ray cluster."
+            )
+
+        # Parse and validate the Azure URI
+        parsed = urlparse(object_uri)
+        scheme = parsed.scheme.lower()
+
+        # Validate URI format: scheme://container@account.domain/path
+        if not parsed.netloc or "@" not in parsed.netloc:
+            raise ValueError(
+                f"Invalid {scheme.upper()} URI format - missing container@account: {object_uri}"
+            )
+
+        container_part, hostname_part = parsed.netloc.split("@", 1)
+
+        # Validate container name (must be non-empty)
+        if not container_part:
+            raise ValueError(
+                f"Invalid {scheme.upper()} URI format - empty container name: {object_uri}"
+            )
+
+        # Validate hostname format based on scheme
+        valid_hostname = False
+        if scheme == "abfss":
+            valid_hostname = hostname_part.endswith(".dfs.core.windows.net")
+            expected_domains = ".dfs.core.windows.net"
+        elif scheme == "azure":
+            valid_hostname = hostname_part.endswith(
+                ".blob.core.windows.net"
+            ) or hostname_part.endswith(".dfs.core.windows.net")
+            expected_domains = ".blob.core.windows.net or .dfs.core.windows.net"
+
+        if not hostname_part or not valid_hostname:
+            raise ValueError(
+                f"Invalid {scheme.upper()} URI format - invalid hostname (must end with {expected_domains}): {object_uri}"
+            )
+
+        # Extract and validate account name
+        azure_storage_account_name = hostname_part.split(".")[0]
+        if not azure_storage_account_name:
+            raise ValueError(
+                f"Invalid {scheme.upper()} URI format - empty account name: {object_uri}"
+            )
+
+        # Create the adlfs filesystem
+        adlfs_fs = adlfs.AzureBlobFileSystem(
+            account_name=azure_storage_account_name,
+            credential=DefaultAzureCredential(),
+        )
+
+        # Wrap with PyArrow's PyFileSystem for compatibility
+        fs = pa_fs.PyFileSystem(pa_fs.FSSpecHandler(adlfs_fs))
+
+        # Return the path without the scheme prefix
+        path = f"{container_part}{parsed.path}"
+
+        return fs, path
+
+    @staticmethod
+    def _create_abfss_filesystem(object_uri: str) -> Tuple[pa_fs.FileSystem, str]:
+        """Create an ABFSS filesystem for Azure Data Lake Storage Gen2.
+
+        This is a wrapper around _create_azure_filesystem for backward compatibility.
+
+        Args:
+            object_uri: ABFSS URI (abfss://container@account.dfs.core.windows.net/path)
+
+        Returns:
+            Tuple of (PyArrow FileSystem, path without abfss:// prefix)
+        """
+        return PyArrowFileSystem._create_azure_filesystem(object_uri)
+
+    @staticmethod
+    def _filter_files(
+        fs: pa_fs.FileSystem,
+        source_path: str,
+        destination_path: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+    ) -> List[Tuple[str, str]]:
+        """Filter files from cloud storage based on inclusion and exclusion criteria.
+
+        Args:
+            fs: PyArrow filesystem instance
+            source_path: Source path in cloud storage
+            destination_path: Local destination path
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude files ending with these suffixes
+
+        Returns:
+            List of tuples containing (source_file_path, destination_file_path)
+        """
+        file_selector = pa_fs.FileSelector(source_path, recursive=True)
+        file_infos = fs.get_file_info(file_selector)
+
+        path_pairs = []
+        for file_info in file_infos:
+            if file_info.type != pa_fs.FileType.File:
+                continue
+
+            rel_path = file_info.path[len(source_path) :].lstrip("/")
+
+            # Apply filters
+            if substrings_to_include:
+                if not any(
+                    substring in rel_path for substring in substrings_to_include
+                ):
+                    continue
+
+            if suffixes_to_exclude:
+                if any(rel_path.endswith(suffix) for suffix in suffixes_to_exclude):
+                    continue
+
+            path_pairs.append(
+                (file_info.path, os.path.join(destination_path, rel_path))
+            )
+
+        return path_pairs
+
+    @staticmethod
+    def get_file(
+        object_uri: str, decode_as_utf_8: bool = True
+    ) -> Optional[Union[str, bytes]]:
+        """Download a file from cloud storage into memory.
+
+        Args:
+            object_uri: URI of the file (s3://, gs://, abfss://, or azure://)
+            decode_as_utf_8: If True, decode the file as UTF-8
+
+        Returns:
+            File contents as string or bytes, or None if file doesn't exist
+        """
+        try:
+            fs, path = PyArrowFileSystem.get_fs_and_path(object_uri)
+
+            # Check if file exists
+            if not fs.get_file_info(path).type == pa_fs.FileType.File:
+                logger.info(f"URI {object_uri} does not exist.")
+                return None
+
+            # Read file
+            with fs.open_input_file(path) as f:
+                body = f.read()
+
+            if decode_as_utf_8:
+                body = body.decode("utf-8")
+            return body
+        except Exception as e:
+            logger.warning(f"Error reading {object_uri}: {e}")
+            return None
+
+    @staticmethod
+    def list_subfolders(folder_uri: str) -> List[str]:
+        """List the immediate subfolders in a cloud directory.
+
+        Args:
+            folder_uri: URI of the directory (s3://, gs://, abfss://, or azure://)
+
+        Returns:
+            List of subfolder names (without trailing slashes)
+        """
+        # Ensure that the folder_uri has a trailing slash.
+        folder_uri = f"{folder_uri.rstrip('/')}/"
+
+        try:
+            fs, path = PyArrowFileSystem.get_fs_and_path(folder_uri)
+
+            # List directory contents
+            file_infos = fs.get_file_info(pa_fs.FileSelector(path, recursive=False))
+
+            # Filter for directories and extract subfolder names
+            subfolders = []
+            for file_info in file_infos:
+                if file_info.type == pa_fs.FileType.Directory:
+                    # Extract just the subfolder name without the full path
+                    subfolder = os.path.basename(file_info.path.rstrip("/"))
+                    subfolders.append(subfolder)
+
+            return subfolders
+        except Exception as e:
+            logger.error(f"Error listing subfolders in {folder_uri}: {e}")
+            return []
+
+    @staticmethod
+    def download_files(
+        path: str,
+        bucket_uri: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+        max_concurrency: int = 10,
+        chunk_size: int = 64 * 1024 * 1024,
+    ) -> None:
+        """Download files from cloud storage to a local directory.
+
+        Args:
+            path: Local directory where files will be downloaded
+            bucket_uri: URI of cloud directory
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude certain files from download
+            max_concurrency: Maximum number of concurrent files to download (default: 10)
+            chunk_size: Size of transfer chunks (default: 64MB)
+        """
+        try:
+            fs, source_path = PyArrowFileSystem.get_fs_and_path(bucket_uri)
+
+            # Ensure destination exists
+            os.makedirs(path, exist_ok=True)
+
+            # If no filters, use direct copy_files
+            if not substrings_to_include and not suffixes_to_exclude:
+                pa_fs.copy_files(
+                    source=source_path,
+                    destination=path,
+                    source_filesystem=fs,
+                    destination_filesystem=pa_fs.LocalFileSystem(),
+                    use_threads=True,
+                    chunk_size=chunk_size,
+                )
+                return
+
+            # List and filter files
+            files_to_download = PyArrowFileSystem._filter_files(
+                fs, source_path, path, substrings_to_include, suffixes_to_exclude
+            )
+
+            if not files_to_download:
+                logger.info("Filters do not match any of the files, skipping download")
+                return
+
+            def download_single_file(file_paths):
+                source_file_path, dest_file_path = file_paths
+                # Create destination directory if needed
+                dest_dir = os.path.dirname(dest_file_path)
+                if dest_dir:
+                    os.makedirs(dest_dir, exist_ok=True)
+
+                # Use PyArrow's copy_files for individual files,
+                pa_fs.copy_files(
+                    source=source_file_path,
+                    destination=dest_file_path,
+                    source_filesystem=fs,
+                    destination_filesystem=pa_fs.LocalFileSystem(),
+                    use_threads=True,
+                    chunk_size=chunk_size,
+                )
+                return dest_file_path
+
+            max_workers = min(max_concurrency, len(files_to_download))
+            with ThreadPoolExecutor(max_workers=max_workers) as executor:
+                futures = [
+                    executor.submit(download_single_file, file_paths)
+                    for file_paths in files_to_download
+                ]
+
+                for future in futures:
+                    try:
+                        future.result()
+                    except Exception as e:
+                        logger.error(f"Failed to download file: {e}")
+                        raise
+
+        except Exception as e:
+            logger.exception(f"Error downloading files from {bucket_uri}: {e}")
+            raise
+
+    @staticmethod
+    def upload_files(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload files to cloud storage.
+
+        Args:
+            local_path: The local path of the files to upload.
+            bucket_uri: The bucket uri to upload the files to, must start with
+                `s3://`, `gs://`, `abfss://`, or `azure://`.
+        """
+        try:
+            fs, dest_path = PyArrowFileSystem.get_fs_and_path(bucket_uri)
+
+            pa_fs.copy_files(
+                source=local_path,
+                destination=dest_path,
+                source_filesystem=pa_fs.LocalFileSystem(),
+                destination_filesystem=fs,
+            )
+        except Exception as e:
+            logger.exception(f"Error uploading files to {bucket_uri}: {e}")
+            raise

python/ray/llm/_internal/common/utils/cloud_filesystem/s3_filesystem.py

@@ -0,0 +1,397 @@
+"""S3-specific filesystem implementation using boto3.
+
+This module provides an S3-specific implementation that uses boto3 (AWS SDK for Python)
+for reliable and efficient S3 operations.
+"""
+
+import os
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from typing import Any, List, Optional, Union
+
+try:
+    import boto3
+    from botocore import UNSIGNED
+    from botocore.config import Config
+except ImportError:
+    boto3 = None  # type: ignore[assignment]
+    UNSIGNED = None  # type: ignore[assignment]
+    Config = None  # type: ignore[assignment]
+
+from ray.llm._internal.common.observability.logging import get_logger
+from ray.llm._internal.common.utils.cloud_filesystem.base import BaseCloudFileSystem
+
+logger = get_logger(__name__)
+
+
+def _check_boto3() -> None:
+    """Raise a clear error if boto3/botocore are not installed."""
+    if boto3 is None:
+        raise ImportError(
+            "boto3 and botocore are required for S3 operations but are not installed. "
+            "Install them with: pip install boto3"
+        )
+
+
+class S3FileSystem(BaseCloudFileSystem):
+    """S3-specific implementation of cloud filesystem operations using boto3.
+
+    This implementation uses boto3 (AWS SDK for Python) for reliable and efficient
+    operations with S3 storage.
+    """
+
+    @staticmethod
+    def _parse_s3_uri(uri: str) -> tuple[str, str, bool]:
+        """Parse S3 URI into bucket and key.
+
+        Args:
+            uri: S3 URI (e.g., s3://bucket/path/to/object or s3://anonymous@bucket/path/to/object)
+
+        Returns:
+            Tuple of (bucket_name, key, is_anonymous)
+
+        Raises:
+            ValueError: If URI is not a valid S3 URI
+        """
+        # Check if anonymous@ prefix exists
+        is_anonymous = False
+        if uri.startswith("s3://anonymous@"):
+            is_anonymous = True
+            uri = uri.replace("s3://anonymous@", "s3://", 1)
+
+        if not uri.startswith("s3://"):
+            raise ValueError(f"Invalid S3 URI: {uri}")
+
+        # Remove s3:// prefix and split into bucket and key
+        path = uri[5:]  # Remove "s3://"
+        parts = path.split("/", 1)
+        bucket = parts[0]
+        key = parts[1] if len(parts) > 1 else ""
+
+        return bucket, key, is_anonymous
+
+    @staticmethod
+    def _get_s3_client(max_pool_connections: int = 50, anonymous: bool = False):
+        """Create a new S3 client instance with connection pooling.
+
+        Args:
+            max_pool_connections: Maximum number of connections in the pool.
+                Should be >= max_workers for optimal performance.
+            anonymous: Whether to use anonymous access to S3.
+
+        Returns:
+            boto3 S3 client with connection pooling configured
+        """
+        _check_boto3()
+        # Configure connection pooling for better concurrent performance
+        config = Config(
+            max_pool_connections=max_pool_connections,
+            # Retry configuration for transient failures
+            retries={
+                "max_attempts": 3,
+                "mode": "adaptive",  # Adapts retry behavior based on error type
+            },
+            # TCP keepalive helps with long-running connections
+            tcp_keepalive=True,
+            signature_version=UNSIGNED if anonymous else None,
+        )
+
+        return boto3.client("s3", config=config)
+
+    @staticmethod
+    def get_file(
+        object_uri: str, decode_as_utf_8: bool = True
+    ) -> Optional[Union[str, bytes]]:
+        """Download a file from cloud storage into memory.
+
+        Args:
+            object_uri: URI of the file (s3://)
+            decode_as_utf_8: If True, decode the file as UTF-8
+
+        Returns:
+            File contents as string or bytes, or None if file doesn't exist
+        """
+        _check_boto3()
+        try:
+            bucket, key, is_anonymous = S3FileSystem._parse_s3_uri(object_uri)
+            s3_client = S3FileSystem._get_s3_client(anonymous=is_anonymous)
+
+            # Download file directly into memory
+            response = s3_client.get_object(Bucket=bucket, Key=key)
+            body = response["Body"].read()
+
+            if decode_as_utf_8:
+                return body.decode("utf-8")
+            return body
+        except Exception as e:
+            logger.error(f"Error reading {object_uri}: {e}")
+
+    @staticmethod
+    def list_subfolders(folder_uri: str) -> List[str]:
+        """List the immediate subfolders in a cloud directory.
+
+        Args:
+            folder_uri: URI of the directory (s3://)
+
+        Returns:
+            List of subfolder names (without trailing slashes)
+        """
+        _check_boto3()
+        try:
+            bucket, prefix, is_anonymous = S3FileSystem._parse_s3_uri(folder_uri)
+
+            # Ensure that the prefix has a trailing slash
+            if prefix and not prefix.endswith("/"):
+                prefix = f"{prefix}/"
+
+            s3_client = S3FileSystem._get_s3_client(anonymous=is_anonymous)
+
+            # List objects with delimiter to get only immediate subfolders
+            response = s3_client.list_objects_v2(
+                Bucket=bucket, Prefix=prefix, Delimiter="/"
+            )
+
+            subfolders = []
+            # CommonPrefixes contains the subdirectories
+            for common_prefix in response.get("CommonPrefixes", []):
+                folder_path = common_prefix["Prefix"]
+                # Extract the folder name from the full prefix
+                # Remove the parent prefix and trailing slash
+                folder_name = folder_path[len(prefix) :].rstrip("/")
+                if folder_name:
+                    subfolders.append(folder_name)
+
+            return subfolders
+        except Exception as e:
+            logger.error(f"Error listing subfolders in {folder_uri}: {e}")
+            return []
+
+    @staticmethod
+    def _calculate_optimal_workers(
+        num_files: int, total_size: int, default_max: int = 100, default_min: int = 10
+    ) -> int:
+        """Calculate optimal number of workers based on file characteristics.
+
+        Args:
+            num_files: Number of files to download
+            total_size: Total size of all files in bytes
+            default_max: Maximum workers to cap at
+            default_min: Minimum workers to use
+
+        Returns:
+            Optimal number of workers between default_min and default_max
+        """
+        if num_files == 0:
+            return default_min
+
+        avg_file_size = total_size / num_files if total_size > 0 else 0
+
+        # Strategy: More workers for smaller files, fewer for larger files
+        if avg_file_size < 1024 * 1024:  # < 1MB (small files)
+            # Use more workers for many small files
+            workers = min(num_files, default_max)
+        elif avg_file_size < 10 * 1024 * 1024:  # 1-10MB (medium files)
+            # Use moderate workers
+            workers = min(num_files // 2, default_max // 2)
+        else:  # > 10MB (large files)
+            # Use fewer workers since each download is bandwidth-intensive
+            workers = min(20, num_files)
+
+        # Ensure workers is between min and max
+        return max(default_min, min(workers, default_max))
+
+    @staticmethod
+    def _download_single_file(
+        s3_client: Any, bucket: str, key: str, local_file_path: str
+    ) -> tuple[str, bool]:
+        """Download a single file from S3.
+
+        Args:
+            s3_client: Shared boto3 S3 client
+            bucket: S3 bucket name
+            key: S3 object key
+            local_file_path: Local path where file will be saved
+
+        Returns:
+            Tuple of (key, success)
+        """
+        try:
+            # Create parent directories if needed
+            os.makedirs(os.path.dirname(local_file_path), exist_ok=True)
+
+            s3_client.download_file(bucket, key, local_file_path)
+            logger.debug(f"Downloaded {key} to {local_file_path}")
+            return key, True
+        except Exception as e:
+            logger.error(f"Failed to download {key}: {e}")
+            return key, False
+
+    @staticmethod
+    def download_files(
+        path: str,
+        bucket_uri: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+        max_workers: Optional[int] = None,
+    ) -> None:
+        """Download files from cloud storage to a local directory concurrently.
+
+        Args:
+            path: Local directory where files will be downloaded
+            bucket_uri: URI of cloud directory
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude certain files from download (e.g .safetensors)
+            max_workers: Maximum number of concurrent downloads. If None, automatically
+                calculated based on file count and sizes (min: 10, max: 100)
+        """
+        try:
+            bucket, prefix, is_anonymous = S3FileSystem._parse_s3_uri(bucket_uri)
+
+            # Ensure the destination directory exists
+            os.makedirs(path, exist_ok=True)
+
+            # Ensure prefix has trailing slash for directory listing
+            if prefix and not prefix.endswith("/"):
+                prefix = f"{prefix}/"
+
+            # Create initial client for listing (will recreate with proper pool size later)
+            s3_client = S3FileSystem._get_s3_client(anonymous=is_anonymous)
+
+            # List all objects in the bucket with the given prefix
+            paginator = s3_client.get_paginator("list_objects_v2")
+            pages = paginator.paginate(Bucket=bucket, Prefix=prefix)
+
+            # Collect all files to download and track total size
+            files_to_download = []
+            total_size = 0
+
+            for page in pages:
+                for obj in page.get("Contents", []):
+                    key = obj["Key"]
+                    size = obj.get("Size", 0)
+
+                    # Skip if it's a directory marker
+                    if key.endswith("/"):
+                        continue
+
+                    # Get the relative path (remove the prefix)
+                    relative_path = key[len(prefix) :]
+
+                    # Apply include filters
+                    if substrings_to_include:
+                        if not any(
+                            substr in relative_path for substr in substrings_to_include
+                        ):
+                            continue
+
+                    # Apply exclude filters
+                    if suffixes_to_exclude:
+                        if any(
+                            relative_path.endswith(suffix.lstrip("*"))
+                            for suffix in suffixes_to_exclude
+                        ):
+                            continue
+
+                    # Construct local file path
+                    local_file_path = os.path.join(path, relative_path)
+                    files_to_download.append((bucket, key, local_file_path))
+                    total_size += size
+
+            # Download files concurrently
+            if not files_to_download:
+                logger.info(f"No files matching filters to download from {bucket_uri}")
+                return
+
+            # Dynamically calculate workers if not provided
+            if max_workers is None:
+                max_workers = S3FileSystem._calculate_optimal_workers(
+                    num_files=len(files_to_download),
+                    total_size=total_size,
+                    default_max=100,
+                    default_min=10,
+                )
+
+            # Create shared client with proper connection pool size for downloads
+            s3_client = S3FileSystem._get_s3_client(
+                max_pool_connections=max_workers + 10, anonymous=is_anonymous
+            )
+
+            failed_downloads = []
+
+            with ThreadPoolExecutor(max_workers=max_workers) as executor:
+                # Submit all download tasks with shared client
+                future_to_key = {
+                    executor.submit(
+                        S3FileSystem._download_single_file,
+                        s3_client,  # Pass shared client to each worker
+                        bucket,
+                        key,
+                        local_path,
+                    ): key
+                    for bucket, key, local_path in files_to_download
+                }
+
+                # Process completed downloads
+                for future in as_completed(future_to_key):
+                    key, success = future.result()
+                    if not success:
+                        failed_downloads.append(key)
+
+            # Report any failures
+            if failed_downloads:
+                logger.error(
+                    f"Failed to download {len(failed_downloads)} files: {failed_downloads[:5]}..."
+                )
+
+        except Exception as e:
+            logger.exception(f"Error downloading files from {bucket_uri}: {e}")
+            raise
+
+    @staticmethod
+    def upload_files(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload files to cloud storage.
+
+        Args:
+            local_path: The local path of the files to upload.
+            bucket_uri: The bucket uri to upload the files to, must start with `s3://`.
+        """
+        try:
+            bucket, prefix, is_anonymous = S3FileSystem._parse_s3_uri(bucket_uri)
+
+            # Ensure prefix has trailing slash for directory upload
+            if prefix and not prefix.endswith("/"):
+                prefix = f"{prefix}/"
+
+            s3_client = S3FileSystem._get_s3_client(anonymous=is_anonymous)
+
+            local_path_obj = Path(local_path)
+
+            # Walk through the local directory and upload each file
+            if local_path_obj.is_file():
+                # Upload a single file
+                file_name = local_path_obj.name
+                s3_key = f"{prefix}{file_name}" if prefix else file_name
+                s3_client.upload_file(str(local_path_obj), bucket, s3_key)
+                logger.debug(f"Uploaded {local_path_obj} to s3://{bucket}/{s3_key}")
+            elif local_path_obj.is_dir():
+                # Upload directory recursively
+                for file_path in local_path_obj.rglob("*"):
+                    if file_path.is_file():
+                        # Calculate relative path from local_path
+                        relative_path = file_path.relative_to(local_path_obj)
+                        # Construct S3 key
+                        s3_key = f"{prefix}{relative_path.as_posix()}"
+                        # Upload file
+                        s3_client.upload_file(str(file_path), bucket, s3_key)
+                        logger.debug(f"Uploaded {file_path} to s3://{bucket}/{s3_key}")
+            else:
+                raise ValueError(
+                    f"Path {local_path} does not exist or is not a file/directory"
+                )
+
+        except Exception as e:
+            logger.exception(f"Error uploading files to {bucket_uri}: {e}")
+            raise

python/ray/llm/_internal/common/utils/cloud_utils.py

@@ -0,0 +1,546 @@
+import asyncio
+import inspect
+import os
+import time
+from pathlib import Path
+from typing import (
+    Any,
+    Awaitable,
+    Callable,
+    Dict,
+    List,
+    NamedTuple,
+    Optional,
+    TypeVar,
+    Union,
+)
+
+from pydantic import Field, field_validator
+
+from ray.llm._internal.common.base_pydantic import BaseModelExtended
+from ray.llm._internal.common.observability.logging import get_logger
+from ray.llm._internal.common.utils.cloud_filesystem import (
+    AzureFileSystem,
+    GCSFileSystem,
+    PyArrowFileSystem,
+    S3FileSystem,
+)
+
+T = TypeVar("T")
+
+logger = get_logger(__name__)
+
+
+def is_remote_path(path: str) -> bool:
+    """Check if the path is a remote path.
+
+    Args:
+        path: The path to check.
+
+    Returns:
+        True if the path is a remote path, False otherwise.
+    """
+    return (
+        path.startswith("s3://")
+        or path.startswith("gs://")
+        or path.startswith("abfss://")
+        or path.startswith("azure://")
+        or path.startswith("pyarrow-")
+    )
+
+
+class ExtraFiles(BaseModelExtended):
+    bucket_uri: str
+    destination_path: str
+
+
+class CloudMirrorConfig(BaseModelExtended):
+    """Unified mirror config for cloud storage (S3, GCS, or Azure).
+
+    Args:
+        bucket_uri: URI of the bucket (s3://, gs://, abfss://, or azure://)
+        extra_files: Additional files to download
+    """
+
+    bucket_uri: Optional[str] = None
+    extra_files: List[ExtraFiles] = Field(default_factory=list)
+
+    @field_validator("bucket_uri")
+    @classmethod
+    def check_uri_format(cls, value):
+        if value is None:
+            return value
+
+        if not is_remote_path(value):
+            raise ValueError(
+                f'Got invalid value "{value}" for bucket_uri. '
+                'Expected a URI that starts with "s3://", "gs://", "abfss://", or "azure://".'
+            )
+        return value
+
+    @property
+    def storage_type(self) -> str:
+        """Returns the storage type ('s3', 'gcs', 'abfss', or 'azure') based on the URI prefix."""
+        if self.bucket_uri is None:
+            return None
+        elif self.bucket_uri.startswith("s3://"):
+            return "s3"
+        elif self.bucket_uri.startswith("gs://"):
+            return "gcs"
+        elif self.bucket_uri.startswith("abfss://"):
+            return "abfss"
+        elif self.bucket_uri.startswith("azure://"):
+            return "azure"
+        return None
+
+
+class LoraMirrorConfig(BaseModelExtended):
+    lora_model_id: str
+    bucket_uri: str
+    max_total_tokens: Optional[int]
+    sync_args: Optional[List[str]] = None
+
+    @field_validator("bucket_uri")
+    @classmethod
+    def check_uri_format(cls, value):
+        if value is None:
+            return value
+
+        if not is_remote_path(value):
+            raise ValueError(
+                f'Got invalid value "{value}" for bucket_uri. '
+                'Expected a URI that starts with "s3://", "gs://", "abfss://", or "azure://".'
+            )
+        return value
+
+    @property
+    def _bucket_name_and_path(self) -> str:
+        for prefix in ["s3://", "gs://", "abfss://", "azure://"]:
+            if self.bucket_uri.startswith(prefix):
+                return self.bucket_uri[len(prefix) :]
+        return self.bucket_uri
+
+    @property
+    def bucket_name(self) -> str:
+        bucket_part = self._bucket_name_and_path.split("/")[0]
+
+        # For ABFSS and Azure URIs, extract container name from container@account format
+        if self.bucket_uri.startswith(("abfss://", "azure://")) and "@" in bucket_part:
+            return bucket_part.split("@")[0]
+
+        return bucket_part
+
+    @property
+    def bucket_path(self) -> str:
+        return "/".join(self._bucket_name_and_path.split("/")[1:])
+
+
+class CloudFileSystem:
+    """A unified interface for cloud file system operations.
+
+    This class provides a simple interface for common operations on cloud storage
+    systems (S3, GCS, Azure) by delegating to provider-specific implementations
+    for optimal performance.
+    """
+
+    @staticmethod
+    def _get_provider_fs(bucket_uri: str):
+        """Get the appropriate provider-specific filesystem class based on URI.
+
+        Args:
+            bucket_uri: URI of the cloud storage (s3://, gs://, abfss://, or azure://)
+
+        Returns:
+            The appropriate filesystem class (S3FileSystem, GCSFileSystem, or AzureFileSystem)
+
+        Raises:
+            ValueError: If the URI scheme is not supported
+        """
+        if bucket_uri.startswith("pyarrow-"):
+            return PyArrowFileSystem
+        elif bucket_uri.startswith("s3://"):
+            return S3FileSystem
+        elif bucket_uri.startswith("gs://"):
+            return GCSFileSystem
+        elif bucket_uri.startswith(("abfss://", "azure://")):
+            return AzureFileSystem
+        else:
+            raise ValueError(f"Unsupported URI scheme: {bucket_uri}")
+
+    @staticmethod
+    def get_file(
+        object_uri: str, decode_as_utf_8: bool = True
+    ) -> Optional[Union[str, bytes]]:
+        """Download a file from cloud storage into memory.
+
+        Args:
+            object_uri: URI of the file (s3://, gs://, abfss://, or azure://)
+            decode_as_utf_8: If True, decode the file as UTF-8
+
+        Returns:
+            File contents as string or bytes, or None if file doesn't exist
+        """
+        fs_class = CloudFileSystem._get_provider_fs(object_uri)
+        return fs_class.get_file(object_uri, decode_as_utf_8)
+
+    @staticmethod
+    def list_subfolders(folder_uri: str) -> List[str]:
+        """List the immediate subfolders in a cloud directory.
+
+        Args:
+            folder_uri: URI of the directory (s3://, gs://, abfss://, or azure://)
+
+        Returns:
+            List of subfolder names (without trailing slashes)
+        """
+        fs_class = CloudFileSystem._get_provider_fs(folder_uri)
+        return fs_class.list_subfolders(folder_uri)
+
+    @staticmethod
+    def download_files(
+        path: str,
+        bucket_uri: str,
+        substrings_to_include: Optional[List[str]] = None,
+        suffixes_to_exclude: Optional[List[str]] = None,
+    ) -> None:
+        """Download files from cloud storage to a local directory.
+
+        Args:
+            path: Local directory where files will be downloaded
+            bucket_uri: URI of cloud directory
+            substrings_to_include: Only include files containing these substrings
+            suffixes_to_exclude: Exclude certain files from download (e.g .safetensors)
+        """
+        fs_class = CloudFileSystem._get_provider_fs(bucket_uri)
+        fs_class.download_files(
+            path, bucket_uri, substrings_to_include, suffixes_to_exclude
+        )
+
+    @staticmethod
+    def download_model(
+        destination_path: str,
+        bucket_uri: str,
+        tokenizer_only: bool,
+        exclude_safetensors: bool = False,
+    ) -> None:
+        """Download a model from cloud storage.
+
+        This downloads a model in the format expected by the HuggingFace transformers
+        library.
+
+        Args:
+            destination_path: Path where the model will be stored
+            bucket_uri: URI of the cloud directory containing the model
+            tokenizer_only: If True, only download tokenizer-related files
+            exclude_safetensors: If True, skip download of safetensor files
+        """
+        try:
+            # Get the provider-specific filesystem
+            fs_class = CloudFileSystem._get_provider_fs(bucket_uri)
+
+            # Construct hash file URI
+            hash_uri = bucket_uri.rstrip("/") + "/hash"
+
+            # Try to download and read hash file
+            hash_content = fs_class.get_file(hash_uri, decode_as_utf_8=True)
+
+            if hash_content is not None:
+                f_hash = hash_content.strip()
+                logger.info(
+                    f"Detected hash file in bucket {bucket_uri}. "
+                    f"Using {f_hash} as the hash."
+                )
+            else:
+                f_hash = "0000000000000000000000000000000000000000"
+                logger.info(
+                    f"Hash file does not exist in bucket {bucket_uri}. "
+                    f"Using default hash {f_hash} - expected behavior - a hash file is optional. "
+                )
+
+            # Write hash to refs/main
+            main_dir = os.path.join(destination_path, "refs")
+            os.makedirs(main_dir, exist_ok=True)
+            with open(os.path.join(main_dir, "main"), "w") as f:
+                f.write(f_hash)
+
+            # Create destination directory
+            destination_dir = os.path.join(destination_path, "snapshots", f_hash)
+            os.makedirs(destination_dir, exist_ok=True)
+
+            logger.info(f'Downloading model files to directory "{destination_dir}".')
+
+            # Download files
+            tokenizer_file_substrings = (
+                ["tokenizer", "config.json", "chat_template"] if tokenizer_only else []
+            )
+
+            safetensors_to_exclude = [".safetensors"] if exclude_safetensors else None
+
+            CloudFileSystem.download_files(
+                path=destination_dir,
+                bucket_uri=bucket_uri,
+                substrings_to_include=tokenizer_file_substrings,
+                suffixes_to_exclude=safetensors_to_exclude,
+            )
+
+        except Exception as e:
+            logger.exception(f"Error downloading model from {bucket_uri}: {e}")
+            raise
+
+    @staticmethod
+    def upload_files(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload files to cloud storage.
+
+        Args:
+            local_path: The local path of the files to upload.
+            bucket_uri: The bucket uri to upload the files to, must start with
+                `s3://`, `gs://`, `abfss://`, or `azure://`.
+        """
+        fs_class = CloudFileSystem._get_provider_fs(bucket_uri)
+        fs_class.upload_files(local_path, bucket_uri)
+
+    @staticmethod
+    def upload_model(
+        local_path: str,
+        bucket_uri: str,
+    ) -> None:
+        """Upload a model to cloud storage.
+
+        Args:
+            local_path: The local path of the model.
+            bucket_uri: The bucket uri to upload the model to, must start with `s3://` or `gs://`.
+        """
+        try:
+            # If refs/main exists, upload as hash, and treat snapshots/<hash> as the model.
+            # Otherwise, this is a custom model, we do not assume folder hierarchy.
+            refs_main = Path(local_path, "refs", "main")
+            if refs_main.exists():
+                model_path = os.path.join(
+                    local_path, "snapshots", refs_main.read_text().strip()
+                )
+                CloudFileSystem.upload_files(
+                    local_path=model_path, bucket_uri=bucket_uri
+                )
+                CloudFileSystem.upload_files(
+                    local_path=str(refs_main),
+                    bucket_uri=os.path.join(bucket_uri, "hash"),
+                )
+            else:
+                CloudFileSystem.upload_files(
+                    local_path=local_path, bucket_uri=bucket_uri
+                )
+            logger.info(f"Uploaded model files to {bucket_uri}.")
+        except Exception as e:
+            logger.exception(f"Error uploading model to {bucket_uri}: {e}")
+            raise
+
+
+class _CacheEntry(NamedTuple):
+    value: Any
+    expire_time: Optional[float]
+
+
+class CloudObjectCache:
+    """A cache that works with both sync and async fetch functions.
+
+    The purpose of this data structure is to cache the result of a function call
+    usually used to fetch a value from a cloud object store.
+
+    The idea is this:
+    - Cloud operations are expensive
+    - In LoRA specifically, we would fetch remote storage to download the model weights
+    at each request.
+    - If the same model is requested many times, we don't want to inflate the time to first token.
+    - We control the cache via not only the least recently used eviction policy, but also
+    by expiring cache entries after a certain time.
+    - If the object is missing, we cache the missing status for a small duration while if
+    the object exists, we cache the object for a longer duration.
+    """
+
+    def __init__(
+        self,
+        max_size: int,
+        fetch_fn: Union[Callable[[str], Any], Callable[[str], Awaitable[Any]]],
+        missing_expire_seconds: Optional[int] = None,
+        exists_expire_seconds: Optional[int] = None,
+        missing_object_value: Any = object(),
+    ):
+        """Initialize the cache.
+
+        Args:
+            max_size: Maximum number of items to store in cache
+            fetch_fn: Function to fetch values (can be sync or async)
+            missing_expire_seconds: How long to cache missing objects (None for no expiration)
+            exists_expire_seconds: How long to cache existing objects (None for no expiration)
+            missing_object_value: Sentinel value used to represent a missing object in the cache.
+        """
+        self._cache: Dict[str, _CacheEntry] = {}
+        self._max_size = max_size
+        self._fetch_fn = fetch_fn
+        self._missing_expire_seconds = missing_expire_seconds
+        self._exists_expire_seconds = exists_expire_seconds
+        self._is_async = inspect.iscoroutinefunction(fetch_fn) or (
+            callable(fetch_fn) and inspect.iscoroutinefunction(fetch_fn.__call__)
+        )
+        self._missing_object_value = missing_object_value
+        # Lock for thread-safe cache access
+        self._lock = asyncio.Lock()
+
+    async def aget(self, key: str) -> Any:
+        """Async get value from cache or fetch it if needed."""
+
+        if not self._is_async:
+            raise ValueError("Cannot use async get() with sync fetch function")
+
+        async with self._lock:
+            value, should_fetch = self._check_cache(key)
+            if not should_fetch:
+                return value
+
+            # Fetch new value
+            value = await self._fetch_fn(key)
+            self._update_cache(key, value)
+            return value
+
+    def get(self, key: str) -> Any:
+        """Sync get value from cache or fetch it if needed."""
+        if self._is_async:
+            raise ValueError("Cannot use sync get() with async fetch function")
+
+        # For sync access, we use a simple check-then-act pattern
+        # This is safe because sync functions are not used in async context
+        value, should_fetch = self._check_cache(key)
+        if not should_fetch:
+            return value
+
+        # Fetch new value
+        value = self._fetch_fn(key)
+        self._update_cache(key, value)
+        return value
+
+    def _check_cache(self, key: str) -> tuple[Any, bool]:
+        """Check if key exists in cache and is valid.
+
+        Args:
+            key: The cache key to check.
+
+        Returns:
+            Tuple of (value, should_fetch)
+            where should_fetch is True if we need to fetch a new value
+        """
+        now = time.monotonic()
+
+        if key in self._cache:
+            value, expire_time = self._cache[key]
+            if expire_time is None or now < expire_time:
+                return value, False
+
+        return None, True
+
+    def _update_cache(self, key: str, value: Any) -> None:
+        """Update cache with new value."""
+        now = time.monotonic()
+
+        # Calculate expiration
+        expire_time = None
+        if (
+            self._missing_expire_seconds is not None
+            or self._exists_expire_seconds is not None
+        ):
+            if value is self._missing_object_value:
+                expire_time = (
+                    now + self._missing_expire_seconds
+                    if self._missing_expire_seconds
+                    else None
+                )
+            else:
+                expire_time = (
+                    now + self._exists_expire_seconds
+                    if self._exists_expire_seconds
+                    else None
+                )
+
+        # Enforce size limit by removing oldest entry if needed
+        # This is an O(n) operation but it's fine since the cache size is usually small.
+        if len(self._cache) >= self._max_size:
+            oldest_key = min(
+                self._cache, key=lambda k: self._cache[k].expire_time or float("inf")
+            )
+            del self._cache[oldest_key]
+
+        self._cache[key] = _CacheEntry(value, expire_time)
+
+    def __len__(self) -> int:
+        return len(self._cache)
+
+
+class CloudModelAccessor:
+    """Unified accessor for models stored in cloud storage (S3 or GCS).
+
+    Args:
+        model_id: The model id to download or upload.
+        mirror_config: The mirror config for the model.
+    """
+
+    def __init__(self, model_id: str, mirror_config: CloudMirrorConfig):
+        self.model_id = model_id
+        self.mirror_config = mirror_config
+
+    def _get_lock_path(self, suffix: str = "") -> Path:
+        return Path(
+            "~", f"{self.model_id.replace('/', '--')}{suffix}.lock"
+        ).expanduser()
+
+    def _get_model_path(self) -> Path:
+        if Path(self.model_id).exists():
+            return Path(self.model_id)
+        # Delayed import to avoid circular dependencies
+        from huggingface_hub.constants import HF_HUB_CACHE
+
+        return Path(
+            HF_HUB_CACHE, f"models--{self.model_id.replace('/', '--')}"
+        ).expanduser()
+
+
+def remote_object_cache(
+    max_size: int,
+    missing_expire_seconds: Optional[int] = None,
+    exists_expire_seconds: Optional[int] = None,
+    missing_object_value: Any = None,
+) -> Callable[[Callable[..., T]], Callable[..., T]]:
+    """A decorator that provides async caching using CloudObjectCache.
+
+    This is a direct replacement for the remote_object_cache/cachetools combination,
+    using CloudObjectCache internally to maintain cache state.
+
+    Args:
+        max_size: Maximum number of items to store in cache
+        missing_expire_seconds: How long to cache missing objects
+        exists_expire_seconds: How long to cache existing objects
+        missing_object_value: Value to use for missing objects
+
+    Returns:
+        A decorator that wraps an async function with cache lookup.
+    """
+
+    def decorator(func: Callable[..., T]) -> Callable[..., T]:
+        # Create a single cache instance for this function
+        cache = CloudObjectCache(
+            max_size=max_size,
+            fetch_fn=func,
+            missing_expire_seconds=missing_expire_seconds,
+            exists_expire_seconds=exists_expire_seconds,
+            missing_object_value=missing_object_value,
+        )
+
+        async def wrapper(*args, **kwargs):
+            # Extract the key from either first positional arg or object_uri kwarg
+            key = args[0] if args else kwargs.get("object_uri")
+            return await cache.aget(key)
+
+        return wrapper
+
+    return decorator

python/ray/llm/_internal/common/utils/download_utils.py

@@ -0,0 +1,324 @@
+import enum
+import os
+from pathlib import Path
+from typing import List, Optional
+
+from filelock import FileLock
+
+from ray.llm._internal.common.callbacks.base import CallbackBase
+from ray.llm._internal.common.observability.logging import get_logger
+from ray.llm._internal.common.utils.cloud_utils import (
+    CloudFileSystem,
+    CloudMirrorConfig,
+    CloudModelAccessor,
+    is_remote_path,
+)
+from ray.llm._internal.common.utils.import_utils import try_import
+
+torch = try_import("torch")
+
+logger = get_logger(__name__)
+
+STREAMING_LOAD_FORMATS = ["runai_streamer", "runai_streamer_sharded", "tensorizer"]
+
+
+class NodeModelDownloadable(enum.Enum):
+    """Defines which files to download from cloud storage."""
+
+    MODEL_AND_TOKENIZER = enum.auto()
+    TOKENIZER_ONLY = enum.auto()
+    EXCLUDE_SAFETENSORS = enum.auto()
+    NONE = enum.auto()
+
+    def __bool__(self):
+        return self != NodeModelDownloadable.NONE
+
+    def union(self, other: "NodeModelDownloadable") -> "NodeModelDownloadable":
+        """Return a NodeModelDownloadable that is a union of this and the other."""
+        if (
+            self == NodeModelDownloadable.MODEL_AND_TOKENIZER
+            or other == NodeModelDownloadable.MODEL_AND_TOKENIZER
+        ):
+            return NodeModelDownloadable.MODEL_AND_TOKENIZER
+        if (
+            self == NodeModelDownloadable.EXCLUDE_SAFETENSORS
+            or other == NodeModelDownloadable.EXCLUDE_SAFETENSORS
+        ):
+            return NodeModelDownloadable.EXCLUDE_SAFETENSORS
+        if (
+            self == NodeModelDownloadable.TOKENIZER_ONLY
+            or other == NodeModelDownloadable.TOKENIZER_ONLY
+        ):
+            return NodeModelDownloadable.TOKENIZER_ONLY
+
+        return NodeModelDownloadable.NONE
+
+
+def get_model_entrypoint(model_id: str) -> str:
+    """Get the path to entrypoint of the model on disk if it exists, otherwise return the model id as is.
+
+    Entrypoint is typically <HF_HUB_CACHE>/models--<model_id>/
+
+    Args:
+        model_id: Hugging Face model ID.
+
+    Returns:
+        The path to the entrypoint of the model on disk if it exists, otherwise the model id as is.
+    """
+    from huggingface_hub.constants import HF_HUB_CACHE
+
+    model_dir = Path(
+        HF_HUB_CACHE, f"models--{model_id.replace('/', '--')}"
+    ).expanduser()
+    if not model_dir.exists():
+        return model_id
+    return str(model_dir.absolute())
+
+
+def get_model_location_on_disk(model_id: str) -> str:
+    """Get the location of the model on disk if exists, otherwise return the model id as is.
+
+    Args:
+        model_id: Hugging Face model ID.
+
+    Returns:
+        The path to the model on disk if it exists, otherwise the model id as is.
+    """
+    model_dir = Path(get_model_entrypoint(model_id))
+    model_id_or_path = model_id
+
+    model_dir_refs_main = Path(model_dir, "refs", "main")
+
+    if model_dir.exists():
+        if model_dir_refs_main.exists():
+            # If refs/main exists, use the snapshot hash to find the model
+            # and check if *config.json (could be config.json for general models
+            # or adapter_config.json for LoRA adapters) exists to make sure it
+            # follows HF model repo structure.
+            with open(model_dir_refs_main, "r") as f:
+                snapshot_hash = f.read().strip()
+
+            snapshot_hash_path = Path(model_dir, "snapshots", snapshot_hash)
+            if snapshot_hash_path.exists() and list(
+                Path(snapshot_hash_path).glob("*config.json")
+            ):
+                model_id_or_path = str(snapshot_hash_path.absolute())
+        else:
+            # If it doesn't have refs/main, it is a custom model repo
+            # and we can just return the model_dir.
+            model_id_or_path = str(model_dir.absolute())
+
+    return model_id_or_path
+
+
+class CloudModelDownloader(CloudModelAccessor):
+    """Unified downloader for models stored in cloud storage (S3 or GCS).
+
+    Args:
+        model_id: The model id to download.
+        mirror_config: The mirror config for the model.
+    """
+
+    def get_model(
+        self,
+        tokenizer_only: bool,
+        exclude_safetensors: bool = False,
+    ) -> str:
+        """Gets a model from cloud storage and stores it locally.
+
+        Args:
+            tokenizer_only: whether to download only the tokenizer files.
+            exclude_safetensors: whether to download safetensors files to disk.
+
+        Returns:
+            File path of model if downloaded, else the model id.
+        """
+        bucket_uri = self.mirror_config.bucket_uri
+
+        if bucket_uri is None:
+            return self.model_id
+
+        # Use different lock paths for different download types to avoid race conditions
+        # where a tokenizer-only download completes and subsequent full model downloads
+        # incorrectly assume the model weights are already cached.
+        if tokenizer_only:
+            lock_suffix = "-tokenizer"
+        elif exclude_safetensors:
+            lock_suffix = "-exclude-safetensors"
+        else:
+            lock_suffix = "-full"
+        lock_path = self._get_lock_path(suffix=lock_suffix)
+        path = self._get_model_path()
+        storage_type = self.mirror_config.storage_type
+
+        try:
+            # Timeout 0 means there will be only one attempt to acquire
+            # the file lock. If it cannot be acquired, a TimeoutError
+            # will be thrown.
+            # This ensures that subsequent processes don't duplicate work.
+            with FileLock(lock_path, timeout=0):
+                try:
+                    if exclude_safetensors:
+                        logger.info("Skipping download of safetensors files.")
+                    CloudFileSystem.download_model(
+                        destination_path=path,
+                        bucket_uri=bucket_uri,
+                        tokenizer_only=tokenizer_only,
+                        exclude_safetensors=exclude_safetensors,
+                    )
+                    logger.info(
+                        "Finished downloading %s for %s from %s storage",
+                        "tokenizer" if tokenizer_only else "model and tokenizer",
+                        self.model_id,
+                        storage_type.upper() if storage_type else "cloud",
+                    )
+                except RuntimeError:
+                    logger.exception(
+                        "Failed to download files for model %s from %s storage",
+                        self.model_id,
+                        storage_type.upper() if storage_type else "cloud",
+                    )
+        except TimeoutError:
+            # If the directory is already locked, then wait but do not do anything.
+            with FileLock(lock_path, timeout=-1):
+                pass
+        return get_model_location_on_disk(self.model_id)
+
+    def get_extra_files(self) -> List[str]:
+        """Gets user-specified extra files from cloud storage and stores them in
+        provided paths.
+
+        Returns: list of file paths of extra files if downloaded.
+        """
+        paths = []
+        extra_files = self.mirror_config.extra_files or []
+        if not extra_files:
+            return paths
+
+        lock_path = self._get_lock_path(suffix="-extra_files")
+        storage_type = self.mirror_config.storage_type
+
+        logger.info(
+            f"Downloading extra files for {self.model_id} from {storage_type} storage"
+        )
+        try:
+            # Timeout 0 means there will be only one attempt to acquire
+            # the file lock. If it cannot be acquired, a TimeoutError
+            # will be thrown.
+            # This ensures that subsequent processes don't duplicate work.
+            with FileLock(lock_path, timeout=0):
+                for extra_file in extra_files:
+                    path = Path(
+                        os.path.expandvars(extra_file.destination_path)
+                    ).expanduser()
+                    paths.append(path)
+                    CloudFileSystem.download_files(
+                        path=path,
+                        bucket_uri=extra_file.bucket_uri,
+                    )
+        except TimeoutError:
+            # If the directory is already locked, then wait but do not do anything.
+            with FileLock(lock_path, timeout=-1):
+                pass
+        return paths
+
+
+def _log_download_info(
+    *, source: str, download_model: NodeModelDownloadable, download_extra_files: bool
+):
+    if download_model == NodeModelDownloadable.NONE:
+        if download_extra_files:
+            logger.info("Downloading extra files from %s", source)
+        else:
+            logger.info("Not downloading anything from %s", source)
+    elif download_model == NodeModelDownloadable.TOKENIZER_ONLY:
+        if download_extra_files:
+            logger.info("Downloading tokenizer and extra files from %s", source)
+        else:
+            logger.info("Downloading tokenizer from %s", source)
+    elif download_model == NodeModelDownloadable.MODEL_AND_TOKENIZER:
+        if download_extra_files:
+            logger.info("Downloading model, tokenizer, and extra files from %s", source)
+        else:
+            logger.info("Downloading model and tokenizer from %s", source)
+
+
+def download_model_files(
+    model_id: Optional[str] = None,
+    mirror_config: Optional[CloudMirrorConfig] = None,
+    download_model: NodeModelDownloadable = NodeModelDownloadable.MODEL_AND_TOKENIZER,
+    download_extra_files: bool = True,
+    callback: Optional[CallbackBase] = None,
+) -> Optional[str]:
+    """
+    Download the model files from the cloud storage. We support two ways to specify
+    the remote model path in the cloud storage:
+    Approach 1:
+    - model_id: The vanilla model id such as "meta-llama/Llama-3.1-8B-Instruct".
+    - mirror_config: Config for downloading model from cloud storage.
+
+    Approach 2:
+    - model_id: The remote path (s3:// or gs://) in the cloud storage.
+    - mirror_config: None.
+    In this approach, we will create a CloudMirrorConfig from the model_id and use that
+    to download the model.
+
+    Args:
+        model_id: The model id.
+        mirror_config: Config for downloading model from cloud storage.
+        download_model: What parts of the model to download.
+        download_extra_files: Whether to download extra files specified in the mirror config.
+        callback: Callback to run before downloading model files.
+
+    Returns:
+        The local path to the downloaded model, or the original model ID
+        if no cloud storage mirror is configured or if the model is not downloaded.
+    """
+
+    # Create the torch cache kernels directory if it doesn't exist.
+    # This is a workaround for a torch issue, where the kernels directory
+    # cannot be created by torch if the parent directory doesn't exist.
+    torch_cache_home = torch.hub._get_torch_home()
+    os.makedirs(os.path.join(torch_cache_home, "kernels"), exist_ok=True)
+    model_path_or_id = model_id
+
+    if callback is not None:
+        callback.run_callback_sync("on_before_download_model_files_distributed")
+
+    if model_id is None:
+        return None
+
+    if mirror_config is None:
+        if is_remote_path(model_id):
+            logger.info(
+                "Creating a CloudMirrorConfig from remote model path %s", model_id
+            )
+            mirror_config = CloudMirrorConfig(bucket_uri=model_id)
+        else:
+            logger.info("No cloud storage mirror configured")
+            return model_id
+
+    storage_type = mirror_config.storage_type
+    source = (
+        f"{storage_type.upper()} mirror" if storage_type else "Cloud storage mirror"
+    )
+
+    _log_download_info(
+        source=source,
+        download_model=download_model,
+        download_extra_files=download_extra_files,
+    )
+
+    downloader = CloudModelDownloader(model_id, mirror_config)
+
+    if download_model != NodeModelDownloadable.NONE:
+        model_path_or_id = downloader.get_model(
+            tokenizer_only=download_model == NodeModelDownloadable.TOKENIZER_ONLY,
+            exclude_safetensors=download_model
+            == NodeModelDownloadable.EXCLUDE_SAFETENSORS,
+        )
+
+    if download_extra_files:
+        downloader.get_extra_files()
+
+    return model_path_or_id

python/ray/llm/_internal/common/utils/import_utils.py

@@ -0,0 +1,90 @@
+"""Utility functions for importing modules in the LLM module."""
+import importlib
+import logging
+from types import ModuleType
+from typing import Any, NoReturn, Optional, Type
+
+logger = logging.getLogger(__name__)
+
+
+def try_import(name: str, error: bool = False) -> Optional[ModuleType]:
+    """Try importing the module and returns the module (or None).
+
+    Args:
+        name: The name of the module to import.
+        error: Whether to raise an error if the module cannot be imported.
+
+    Returns:
+        The module, or None if it cannot be imported.
+
+    Raises:
+        ImportError: If error=True and the module is not installed.
+    """
+    try:
+        return importlib.import_module(name)
+    except ImportError:
+        if error:
+            raise ImportError(f"Could not import {name}")
+        else:
+            logger.warning("Could not import %s", name)
+    return None
+
+
+def raise_llm_engine_import_error(
+    vllm_error: ImportError,
+    sglang_error: ImportError,
+) -> NoReturn:
+    """Raise a descriptive ImportError when both vLLM and SGLang fail to import.
+
+    Distinguishes between a package not being installed (ModuleNotFoundError
+    whose .name matches the top-level package) and a broken installation
+    (any other ImportError, e.g. a missing .so or a missing transitive dep).
+
+    Args:
+        vllm_error: The ImportError raised when importing vLLM.
+        sglang_error: The ImportError raised when importing SGLang.
+    """
+    vllm_not_installed = (
+        isinstance(vllm_error, ModuleNotFoundError) and vllm_error.name == "vllm"
+    )
+    sglang_not_installed = (
+        isinstance(sglang_error, ModuleNotFoundError) and sglang_error.name == "sglang"
+    )
+
+    if vllm_not_installed and sglang_not_installed:
+        raise ImportError(
+            "Neither vLLM nor SGLang is installed. At least one is required "
+            "for Ray Serve LLM protocol models. Install with: "
+            "`pip install ray[llm]` or `pip install sglang[all]`"
+        )
+
+    messages = []
+    if not vllm_not_installed:
+        messages.append(
+            "vLLM is installed but failed to import. This may indicate a "
+            "CUDA version mismatch or a missing vLLM dependency. "
+            f"Original error: {vllm_error}"
+        )
+    if not sglang_not_installed:
+        messages.append(
+            "SGLang is installed but failed to import. This may indicate a "
+            "missing SGLang dependency. "
+            f"Original error: {sglang_error}"
+        )
+    # Chain to the error that is actually relevant: vLLM's if it is broken,
+    # otherwise sglang's (i.e. vLLM was simply not installed).
+    cause = vllm_error if not vllm_not_installed else sglang_error
+    raise ImportError("\n".join(messages)) from cause
+
+
+def load_class(path: str) -> Type[Any]:
+    """Load class from string path."""
+    if ":" in path:
+        module_path, class_name = path.rsplit(":", 1)
+    else:
+        module_path, class_name = path.rsplit(".", 1)
+
+    module = try_import(module_path, error=True)
+    callback_class = getattr(module, class_name)
+
+    return callback_class

python/ray/llm/_internal/common/utils/lora_utils.py

@@ -0,0 +1,233 @@
+"""
+Generic LoRA utilities and abstractions.
+
+This module provides canonical LoRA utility functions for both serve and batch components.
+It serves as the single source of truth for LoRA operations and builds on the generic
+download primitives from download_utils.py.
+"""
+
+import json
+import os
+import subprocess
+import time
+from functools import wraps
+from typing import Any, Callable, List, Optional, TypeVar, Union
+
+from ray.llm._internal.common.constants import (
+    CLOUD_OBJECT_EXISTS_EXPIRE_S,
+    CLOUD_OBJECT_MISSING_EXPIRE_S,
+    LORA_ADAPTER_CONFIG_NAME,
+)
+
+# Import the global ID manager from common models
+from ray.llm._internal.common.models import make_async
+from ray.llm._internal.common.observability.logging import get_logger
+from ray.llm._internal.common.utils.cloud_utils import (
+    CloudFileSystem,
+    is_remote_path,
+    remote_object_cache,
+)
+from ray.llm._internal.common.utils.download_utils import (
+    CloudMirrorConfig,
+    CloudModelDownloader,
+)
+
+logger = get_logger(__name__)
+
+# Sentinel object for missing cloud objects
+CLOUD_OBJECT_MISSING = object()
+
+DEFAULT_LORA_MAX_TOTAL_TOKENS = 4096
+T = TypeVar("T")
+
+
+def get_base_model_id(model_id: str) -> str:
+    """Get base model id for a given model id."""
+    return model_id.split(":")[0]
+
+
+def get_lora_id(lora_model_id: str) -> str:
+    """Get lora id for a given lora model id."""
+    return ":".join(lora_model_id.split(":")[1:])
+
+
+def clean_model_id(model_id: str) -> str:
+    """Clean model ID for filesystem usage by replacing slashes with dashes."""
+    return model_id.replace("/", "--")
+
+
+def clear_directory(dir: str) -> None:
+    """Clear a directory recursively, ignoring missing directories."""
+    try:
+        subprocess.run(f"rm -r {dir}", shell=True, check=False)
+    except FileNotFoundError:
+        pass
+
+
+def retry_with_exponential_backoff(
+    max_tries: int,
+    exception_to_check: type[Exception],
+    base_delay: float = 1,
+    max_delay: float = 32,
+    exponential_base: float = 2,
+) -> Callable[[Callable[..., T]], Callable[..., T]]:
+    """Retry decorator with exponential backoff."""
+
+    def decorator(func: Callable[..., T]) -> Callable[..., T]:
+        @wraps(func)
+        def wrapper(*args: Any, **kwargs: Any) -> T:
+            delay = base_delay
+            last_exception = None
+
+            for attempt in range(max_tries):
+                try:
+                    return func(*args, **kwargs)
+                except exception_to_check as e:
+                    last_exception = e
+                    if attempt == max_tries - 1:  # Last attempt
+                        raise last_exception
+
+                    # Log the failure and retry
+                    logger.warning(
+                        f"Attempt {attempt + 1}/{max_tries} failed: {str(e)}. "
+                        f"Retrying in {delay} seconds..."
+                    )
+                    time.sleep(delay)
+                    # Calculate next delay with exponential backoff
+                    delay = min(delay * exponential_base, max_delay)
+
+            # This should never be reached due to the raise in the loop
+            raise last_exception if last_exception else RuntimeError(
+                "Unexpected error in retry logic"
+            )
+
+        return wrapper
+
+    return decorator
+
+
+def sync_files_with_lock(
+    bucket_uri: str,
+    local_path: str,
+    timeout: Optional[float] = None,
+    substrings_to_include: Optional[List[str]] = None,
+) -> None:
+    """Sync files from bucket_uri to local_path with file locking."""
+    from filelock import FileLock
+
+    logger.info("Downloading %s to %s", bucket_uri, local_path)
+
+    with FileLock(local_path + ".lock", timeout=timeout or -1):
+        try:
+            CloudFileSystem.download_files(
+                path=local_path,
+                bucket_uri=bucket_uri,
+                substrings_to_include=substrings_to_include,
+            )
+        except Exception as e:
+            logger.error(
+                "Failed to sync files from %s to %s: %s",
+                bucket_uri,
+                local_path,
+                str(e),
+            )
+            raise
+
+
+@make_async
+def _get_object_from_cloud(object_uri: str) -> Union[str, object]:
+    """Gets an object from the cloud."""
+    if object_uri.endswith("/"):
+        raise ValueError(f'object_uri {object_uri} must not end with a "/".')
+
+    body_str = CloudFileSystem.get_file(object_uri)
+
+    if body_str is None:
+        logger.info(f"{object_uri} does not exist.")
+        return CLOUD_OBJECT_MISSING
+    else:
+        return body_str
+
+
+@remote_object_cache(
+    max_size=4096,
+    missing_expire_seconds=CLOUD_OBJECT_MISSING_EXPIRE_S,
+    exists_expire_seconds=CLOUD_OBJECT_EXISTS_EXPIRE_S,
+    missing_object_value=CLOUD_OBJECT_MISSING,
+)
+async def get_object_from_cloud(object_uri: str) -> Union[str, object]:
+    """Gets an object from the cloud with caching."""
+    return await _get_object_from_cloud(object_uri)
+
+
+async def get_lora_finetuned_context_length(bucket_uri: str) -> Optional[int]:
+    """Gets the sequence length used to tune the LoRA adapter."""
+    if bucket_uri.endswith("/"):
+        bucket_uri = bucket_uri.rstrip("/")
+    object_uri = f"{bucket_uri}/{LORA_ADAPTER_CONFIG_NAME}"
+
+    object_str_or_missing_message = await get_object_from_cloud(object_uri)
+
+    if object_str_or_missing_message is CLOUD_OBJECT_MISSING:
+        logger.debug(f"LoRA adapter config file not found at {object_uri}")
+        return None
+
+    try:
+        adapter_config_str = object_str_or_missing_message
+        adapter_config = json.loads(adapter_config_str)
+        return adapter_config.get("max_length")
+    except (json.JSONDecodeError, AttributeError) as e:
+        logger.warning(f"Failed to parse LoRA adapter config at {object_uri}: {e}")
+        return None
+
+
+def get_lora_model_ids(
+    dynamic_lora_loading_path: str,
+    base_model_id: str,
+) -> List[str]:
+    """Get the model IDs of all the LoRA models.
+
+    The dynamic_lora_loading_path is expected to hold subfolders each for
+    a different lora checkpoint. Each subfolder name will correspond to
+    the unique identifier for the lora checkpoint. The lora model is
+    accessible via <base_model_id>:<lora_id>. Therefore, we prepend
+    the base_model_id to each subfolder name.
+
+    Args:
+        dynamic_lora_loading_path: the cloud folder that contains all the LoRA
+            weights.
+        base_model_id: model ID of the base model.
+
+    Returns:
+        List of LoRA fine-tuned model IDs. Does not include the base model
+        itself.
+    """
+    lora_subfolders = CloudFileSystem.list_subfolders(dynamic_lora_loading_path)
+
+    lora_model_ids = []
+    for subfolder in lora_subfolders:
+        lora_model_ids.append(f"{base_model_id}:{subfolder}")
+
+    return lora_model_ids
+
+
+def download_lora_adapter(
+    lora_name: str,
+    remote_path: Optional[str] = None,
+) -> str:
+    """Download a LoRA adapter from remote storage.
+
+    This maintains backward compatibility with existing code.
+    """
+
+    assert not is_remote_path(
+        lora_name
+    ), "lora_name cannot be a remote path (s3:// or gs://)"
+
+    if remote_path is None:
+        return lora_name
+
+    lora_path = os.path.join(remote_path, lora_name)
+    mirror_config = CloudMirrorConfig(bucket_uri=lora_path)
+    downloader = CloudModelDownloader(lora_name, mirror_config)
+    return downloader.get_model(tokenizer_only=False)

python/ray/llm/_internal/common/utils/upload_utils.py

@@ -0,0 +1,126 @@
+from pathlib import Path
+
+import typer
+from filelock import FileLock
+from typing_extensions import Annotated
+
+from ray.llm._internal.common.observability.logging import get_logger
+from ray.llm._internal.common.utils.cloud_utils import (
+    CloudFileSystem,
+    CloudMirrorConfig,
+    CloudModelAccessor,
+    is_remote_path,
+)
+from ray.llm._internal.common.utils.download_utils import (
+    get_model_entrypoint,
+)
+
+logger = get_logger(__name__)
+
+
+class CloudModelUploader(CloudModelAccessor):
+    """Unified uploader to upload models to cloud storage (S3 or GCS).
+
+    Args:
+        model_id: The model id to upload.
+        mirror_config: The mirror config for the model.
+    """
+
+    def upload_model(self) -> str:
+        """Upload the model to cloud storage (s3 or gcs).
+
+        Returns:
+            The remote path of the uploaded model.
+        """
+        bucket_uri = self.mirror_config.bucket_uri
+
+        lock_path = self._get_lock_path()
+        path = self._get_model_path()
+        storage_type = self.mirror_config.storage_type
+
+        try:
+            # Timeout 0 means there will be only one attempt to acquire
+            # the file lock. If it cannot be acquired, a TimeoutError
+            # will be thrown.
+            # This ensures that subsequent processes don't duplicate work.
+            with FileLock(lock_path, timeout=0):
+                try:
+                    CloudFileSystem.upload_model(
+                        local_path=path,
+                        bucket_uri=bucket_uri,
+                    )
+                    logger.info(
+                        "Finished uploading %s to %s storage",
+                        self.model_id,
+                        storage_type.upper() if storage_type else "cloud",
+                    )
+                except RuntimeError:
+                    logger.exception(
+                        "Failed to upload model %s to %s storage",
+                        self.model_id,
+                        storage_type.upper() if storage_type else "cloud",
+                    )
+        except TimeoutError:
+            # If the directory is already locked, then wait but do not do anything.
+            with FileLock(lock_path, timeout=-1):
+                pass
+        return bucket_uri
+
+
+def upload_model_files(model_id: str, bucket_uri: str) -> str:
+    """Upload the model files to cloud storage (s3 or gcs).
+
+    If `model_id` is a local path, the files will be uploaded to the cloud storage.
+    If `model_id` is a huggingface model id, the model will be downloaded from huggingface
+    and then uploaded to the cloud storage.
+
+    Args:
+        model_id: The huggingface model id, or local model path to upload.
+        bucket_uri: The bucket uri to upload the model to, must start with `s3://` or `gs://`.
+
+    Returns:
+        The remote path of the uploaded model.
+    """
+    assert not is_remote_path(
+        model_id
+    ), f"model_id must NOT be a remote path: {model_id}"
+    assert is_remote_path(bucket_uri), f"bucket_uri must be a remote path: {bucket_uri}"
+
+    if not Path(model_id).exists():
+        maybe_downloaded_model_path = get_model_entrypoint(model_id)
+        if not Path(maybe_downloaded_model_path).exists():
+            logger.info(
+                "Assuming %s is huggingface model id, and downloading it.", model_id
+            )
+            import huggingface_hub
+
+            huggingface_hub.snapshot_download(repo_id=model_id)
+            # Try to get the model path again after downloading.
+            maybe_downloaded_model_path = get_model_entrypoint(model_id)
+            assert Path(
+                maybe_downloaded_model_path
+            ).exists(), f"Failed to download the model {model_id} to {maybe_downloaded_model_path}"
+            return upload_model_files(maybe_downloaded_model_path, bucket_uri)
+        else:
+            return upload_model_files(maybe_downloaded_model_path, bucket_uri)
+
+    uploader = CloudModelUploader(model_id, CloudMirrorConfig(bucket_uri=bucket_uri))
+    return uploader.upload_model()
+
+
+def upload_model_cli(
+    model_source: Annotated[
+        str,
+        typer.Option(
+            help="HuggingFace model ID to download, or local model path to upload",
+        ),
+    ],
+    bucket_uri: Annotated[
+        str,
+        typer.Option(
+            help="The bucket uri to upload the model to, must start with `s3://` or `gs://`",
+        ),
+    ],
+):
+    """Upload the model files to cloud storage (s3 or gcs)."""
+    upload_model_files(model_source, bucket_uri)