Fix content length mismatch

litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

@@ -402,13 +402,14 @@ async def non_streaming_http_request_handler(
                 requested_query_params=requested_query_params,
             )
         else:
+            json_str = json.dumps(_parsed_body)  # Pre-serialize JSON to avoid Content-Length mismatch
             # Generic httpx method
             response = await async_client.request(
                 method=request.method,
                 url=url,
                 headers=headers,
                 params=requested_query_params,
-                json=_parsed_body,
+                content=json_str.encode()
             )
         return response
 
@@ -584,10 +585,11 @@ async def pass_through_request(  # noqa: PLR0915
             },
         )
         if stream:
+            json_str = json.dumps(_parsed_body)  # Pre-serialize JSON to avoid Content-Length mismatch
             req = async_client.build_request(
                 "POST",
                 url,
-                json=_parsed_body,
+                content=json_str.encode(),
                 params=requested_query_params,
                 headers=headers,
             )

tests/litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py

@@ -36,6 +36,112 @@ def test_is_multipart():
     assert HttpPassThroughEndpointHelpers.is_multipart(request) is False
 
 
+# Test content length consistency for pass-through endpoints
+def test_content_length_consistency():
+    """Test that the Content-Length is consistent when using pre-serialized JSON."""
+    # Test data
+    test_data = {
+        "prompt": "\n\nHuman: Tell me a short joke\n\nAssistant:",
+        "max_tokens_to_sample": 50,
+        "temperature": 0.7,
+        "top_p": 0.9
+    }
+
+    # Method 1: Using json parameter (what causes the issue)
+    request1 = httpx.Request(
+        method="POST",
+        url="https://example.com",
+        json=test_data
+    )
+
+    # Method 2: Using data parameter with pre-serialized JSON (our fix)
+    json_str = json.dumps(test_data)
+    request2 = httpx.Request(
+        method="POST",
+        url="https://example.com",
+        content=json_str.encode(),
+        headers={"Content-Type": "application/json"}
+    )
+
+    # Print the actual differences for verification
+    print(f"Method 1 (json): Content-Length={request1.headers.get('content-length')}, Actual={len(request1.content)}")
+    print(f"Method 2 (data): Content-Length={request2.headers.get('content-length')}, Actual={len(request2.content)}")
+    print(f"Method 1 body: {request1.content}")
+    print(f"Method 2 body: {request2.content}")
+
+    # Assert that the Content-Length header matches the actual body length for our fix
+    assert len(request2.content) == int(request2.headers.get("content-length", 0))
+
+    # Demonstrate the potential mismatch with the json parameter
+    # Note: This might not always fail depending on how httpx serializes JSON,
+    # but it demonstrates the potential issue
+    json_str_manual = json.dumps(test_data)
+    assert len(json_str_manual.encode()) != len(request1.content), "JSON serialization should be different"
+
+
+@pytest.mark.parametrize("use_data", [True, False])
+def test_aws_sigv4_content_length_consistency(use_data):
+    """
+    Test that demonstrates how using data with pre-serialized JSON ensures
+    Content-Length consistency for AWS SigV4 authentication.
+    """
+    # Test data
+    test_data = {
+        "prompt": "\n\nHuman: Tell me a short joke\n\nAssistant:",
+        "max_tokens_to_sample": 50,
+        "temperature": 0.7,
+        "top_p": 0.9
+    }
+
+    # Simulate SigV4 authentication process
+    # 1. Pre-serialize JSON for signing
+    json_str = json.dumps(test_data)
+    content_length_for_signing = len(json_str.encode())
+
+    # 2. Create the actual request
+    if use_data:
+        # Our fix: Use pre-serialized JSON with data parameter
+        request = httpx.Request(
+            method="POST",
+            url="https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-v2/invoke",
+            content=json_str.encode(),
+            headers={
+                "Content-Type": "application/json",
+                "Content-Length": str(content_length_for_signing)
+            }
+        )
+    else:
+        # Original approach: Use json parameter (which causes the issue)
+        request = httpx.Request(
+            method="POST",
+            url="https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-v2/invoke",
+            json=test_data,
+            headers={
+                "Content-Type": "application/json",
+                "Content-Length": str(content_length_for_signing)
+            }
+        )
+
+    # Check if Content-Length matches actual content length
+    actual_content_length = len(request.content)
+    expected_content_length = int(request.headers.get("content-length", 0))
+
+    print(f"Use data: {use_data}")
+    print(f"Expected Content-Length: {expected_content_length}")
+    print(f"Actual content length: {actual_content_length}")
+    print(f"Content: {request.content}")
+
+    if use_data:
+        # Our fix should ensure Content-Length matches
+        assert actual_content_length == expected_content_length, "Content-Length mismatch with data parameter"
+    else:
+        # The original approach might cause a mismatch
+        # Note: This might not always fail depending on how httpx serializes JSON
+        if actual_content_length != expected_content_length:
+            print("Content-Length mismatch detected with json parameter!")
+            print(f"This demonstrates the issue fixed by our PR.")
+
+
 # Test _build_request_files_from_upload_file
 @pytest.mark.asyncio
 async def test_build_request_files_from_upload_file():