Fix overiding of fastapi_response headers

litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

@@ -344,7 +344,13 @@ class HttpPassThroughEndpointHelpers(BasePassthroughUtils):
         if litellm_call_id:
             return_headers["x-litellm-call-id"] = litellm_call_id
         if custom_headers:
-            return_headers.update(custom_headers)
+            # Ensure custom headers don't override actual upstream response headers or let framework defaults (like content-length: 0) interfere.
+            sanitized_custom_headers = {
+                key: value
+                for key, value in custom_headers.items()
+                if key.lower() not in excluded_headers
+            }
+            return_headers.update(sanitized_custom_headers)
 
         return return_headers
 

tests/pass_through_unit_tests/test_pass_through_unit_tests.py

@@ -114,6 +114,40 @@ def test_update_metadata_with_tags_in_header_with_tags(mock_request):
     assert result == {"existing": "value", "tags": ["tag1", "tag2", "tag3"]}
 
 
+def test_get_response_headers_filters_excluded_custom_headers():
+    """
+    Regression test:
+    Ensure excluded headers from FastAPI defaults (e.g. content-length: 0)
+    do not override passthrough response headers.
+    """
+    upstream_headers = httpx.Headers(
+        {
+            "content-type": "application/json",
+            "x-amzn-requestid": "req-123",
+            "content-length": "999",  # should be excluded
+        }
+    )
+
+    custom_headers = {
+        "x-litellm-version": "1.84.0",
+        "content-length": "0",  # should be excluded
+        "server": "uvicorn",  # should be excluded
+    }
+
+    result = HttpPassThroughEndpointHelpers.get_response_headers(
+        headers=upstream_headers,
+        litellm_call_id="call-123",
+        custom_headers=custom_headers,
+    )
+
+    assert result["content-type"] == "application/json"
+    assert result["x-amzn-requestid"] == "req-123"
+    assert result["x-litellm-version"] == "1.84.0"
+    assert result["x-litellm-call-id"] == "call-123"
+    assert "content-length" not in result
+    assert "server" not in result
+
+
 def test_init_kwargs_for_pass_through_endpoint_basic(
     mock_request, mock_user_api_key_dict
 ):