fix: duck-type IO check for Python 3.9/3.10 SpooledTemporaryFile compat; fix zero-byte file_size guard

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -2,12 +2,12 @@
 Common utility functions used for translating messages across providers
 """
 
-import io
 import mimetypes
 import re
 from os import PathLike
 from pathlib import Path
 from typing import (
+    IO,
     TYPE_CHECKING,
     Any,
     Dict,
@@ -718,19 +718,19 @@ def extract_file_data(file_data: FileTypes) -> ExtractedFileData:
                 filename = Path(str(file_content)).name
         with open(file_content, "rb") as f:
             content = f.read()
-    elif isinstance(file_content, io.IOBase):
-        # If it's a file-like object, keep it as-is to avoid loading the entire
-        # file into memory.  Callers that need bytes can call .read() themselves;
-        # callers streaming the data to an HTTP request should pass the object
-        # directly so the transfer is chunked.
-        if not filename and hasattr(file_content, "name"):
-            filename = Path(file_content.name).name
+    elif hasattr(file_content, "read") and hasattr(file_content, "seek"):
+        # Duck-type check covers io.IOBase subclasses AND SpooledTemporaryFile on
+        # Python < 3.11 (which does not inherit from io.IOBase but is still
+        # file-like).  Keep as-is to avoid loading the entire file into memory.
+        file_like = cast(IO[bytes], file_content)
+        if not filename and hasattr(file_like, "name"):
+            filename = Path(file_like.name).name
 
         # Compute file size via seek/tell so providers that need Content-Length
         # (e.g. Gemini resumable upload) don't have to load all bytes.
-        file_content.seek(0, 2)
-        file_size: int = file_content.tell()
-        file_content.seek(0)
+        file_like.seek(0, 2)
+        file_size: int = file_like.tell()
+        file_like.seek(0)
 
         return ExtractedFileData(
             filename=filename,

litellm/llms/custom_httpx/llm_http_handler.py

@@ -2986,7 +2986,7 @@ class BaseLLMHTTPHandler:
                     "headers": transformed_request["upload_request"]["headers"],
                     "timeout": timeout,
                 }
-                if isinstance(upload_data, io.IOBase):
+                if hasattr(upload_data, "read") and hasattr(upload_data, "seek"):
                     upload_kwargs["content"] = upload_data
                 else:
                     upload_kwargs["data"] = upload_data
@@ -3023,7 +3023,9 @@ class BaseLLMHTTPHandler:
         ):
             # Handle traditional file uploads (str, bytes, or IO[bytes] for streaming)
             http_method = provider_config.file_upload_http_method.upper()
-            if isinstance(transformed_request, io.IOBase):
+            if hasattr(transformed_request, "read") and hasattr(
+                transformed_request, "seek"
+            ):
                 # Stream the IO object without loading it fully into memory
                 upload_kwargs: Dict[str, Any] = {
                     "url": api_base,
@@ -3154,7 +3156,7 @@ class BaseLLMHTTPHandler:
                     "headers": transformed_request["upload_request"]["headers"],
                     "timeout": timeout,
                 }
-                if isinstance(upload_data, io.IOBase):
+                if hasattr(upload_data, "read") and hasattr(upload_data, "seek"):
                     async_upload_kwargs["content"] = upload_data
                 else:
                     async_upload_kwargs["data"] = upload_data
@@ -3192,7 +3194,9 @@ class BaseLLMHTTPHandler:
         ):
             # Handle traditional file uploads (str, bytes, or IO[bytes] for streaming)
             http_method = provider_config.file_upload_http_method.upper()
-            if isinstance(transformed_request, io.IOBase):
+            if hasattr(transformed_request, "read") and hasattr(
+                transformed_request, "seek"
+            ):
                 # Stream the IO object without loading it fully into memory
                 async_upload_kwargs: Dict[str, Any] = {
                     "url": api_base,

litellm/llms/gemini/files/transformation.py

@@ -126,7 +126,13 @@ class GoogleAIStudioFilesHandler(GeminiModelInfo, BaseFilesConfig):
 
         # Get file size — prefer the pre-computed value (available when content is
         # an IO[bytes] object so we didn't have to load all bytes into memory).
-        file_size = extracted_data.get("file_size") or len(extracted_data["content"])  # type: ignore[arg-type]
+        # Use explicit `is not None` to avoid treating a 0-byte file as falsy.
+        _precomputed = extracted_data.get("file_size")
+        file_size = (
+            _precomputed
+            if _precomputed is not None
+            else len(extracted_data["content"])  # type: ignore[arg-type]
+        )
 
         # Step 1: Initial resumable upload request
         headers = {

litellm/llms/vertex_ai/files/transformation.py

@@ -1,4 +1,3 @@
-import io
 import json
 import os
 import time
@@ -293,7 +292,9 @@ class VertexAIFilesConfig(VertexBase, BaseFilesConfig):
             return "\n".join(json.dumps(item) for item in vertex_jsonl_content)
         elif isinstance(extracted_file_data_content, bytes):
             return extracted_file_data_content
-        elif isinstance(extracted_file_data_content, io.IOBase):
+        elif hasattr(extracted_file_data_content, "read") and hasattr(
+            extracted_file_data_content, "seek"
+        ):
             return extracted_file_data_content  # type: ignore[return-value]
         else:
             raise ValueError("Unsupported file content type")