fix(callbacks): await async input hooks before provider dispatch

litellm/utils.py

@@ -2078,6 +2078,16 @@ def client(original_function):
                 and _caching_handler_response.embedding_uncached_input is not None
                 else kwargs
             )
+            # Sync integrations still run in Logging.pre_call inside the provider.
+            # Async input callbacks need this event loop; reject before dispatch.
+            if litellm._async_input_callback:
+                details = logging_obj.model_call_details
+                details["model"] = model
+                details["messages"] = logging_obj.messages
+                details["log_event_type"] = "pre_api_call"
+                for callback in tuple(litellm._async_input_callback):
+                    if callable(callback):
+                        await callback(details)
             try:
                 result = await original_function(*args, **call_kwargs)
             except Exception as deployment_error:

tests/unit/test_async_input_callback_dispatch.py

@@ -0,0 +1,90 @@
+"""Async input callbacks must run before an async request reaches the provider."""
+
+import pytest
+
+import litellm
+from litellm import main as litellm_main
+
+
+@pytest.mark.asyncio
+async def test_async_input_callback_runs_once_with_sync_callbacks(monkeypatch):
+    seen = []
+
+    def sync_callback(details):
+        seen.append(("sync", details["model"]))
+
+    async def async_callback(details):
+        seen.append(("async", details["model"]))
+
+    monkeypatch.setattr(litellm, "input_callback", [])
+    monkeypatch.setattr(litellm, "_async_input_callback", [])
+    litellm.logging_callback_manager.add_litellm_input_callback(sync_callback)
+    litellm.logging_callback_manager.add_litellm_input_callback(async_callback)
+    original_provider = litellm_main.mock_completion
+
+    def fake_provider(*args, **kwargs):
+        seen.append(("provider", kwargs.get("model", args[0] if args else "gpt-4o-mini")))
+        return original_provider(*args, **kwargs)
+
+    monkeypatch.setattr(litellm_main, "mock_completion", fake_provider)
+    await litellm.acompletion(model="gpt-4o-mini", messages=[{"role": "user", "content": "hi"}], mock_response="ok")
+    assert seen == [("async", "gpt-4o-mini"), ("provider", "gpt-4o-mini"), ("sync", "gpt-4o-mini")]
+
+
+@pytest.mark.asyncio
+async def test_async_input_callback_error_blocks_mock_provider(monkeypatch):
+    seen = []
+
+    def sync_callback(details):
+        seen.append("sync")
+
+    async def reject(details):
+        seen.append("async")
+        raise RuntimeError("input denied")
+
+    monkeypatch.setattr(litellm, "input_callback", [])
+    monkeypatch.setattr(litellm, "_async_input_callback", [])
+    litellm.logging_callback_manager.add_litellm_input_callback(sync_callback)
+    litellm.logging_callback_manager.add_litellm_input_callback(reject)
+
+    def provider(*args, **kwargs):
+        seen.append("provider")
+        raise AssertionError("provider must not be called")
+
+    monkeypatch.setattr(litellm_main, "mock_completion", provider)
+    with pytest.raises(Exception, match="input denied"):
+        await litellm.acompletion(
+            model="gpt-4o-mini", messages=[{"role": "user", "content": "hi"}], mock_response="ok", num_retries=0
+        )
+    assert seen == ["async"]
+
+
+def test_sync_completion_never_runs_async_input_callback(monkeypatch):
+    seen = []
+
+    async def async_callback(details):
+        seen.append("async")
+
+    def sync_callback(details):
+        seen.append("sync")
+
+    monkeypatch.setattr(litellm, "input_callback", [])
+    monkeypatch.setattr(litellm, "_async_input_callback", [])
+    litellm.logging_callback_manager.add_litellm_input_callback(async_callback)
+    litellm.logging_callback_manager.add_litellm_input_callback(sync_callback)
+    litellm.completion(model="gpt-4o-mini", messages=[{"role": "user", "content": "hi"}], mock_response="ok")
+    assert seen == ["sync"]
+
+
+@pytest.mark.asyncio
+async def test_direct_async_input_registration_runs_once(monkeypatch):
+    seen = []
+
+    async def async_callback(details):
+        seen.append(details["model"])
+
+    monkeypatch.setattr(litellm, "input_callback", [async_callback])
+    monkeypatch.setattr(litellm, "_async_input_callback", [])
+    await litellm.acompletion(model="gpt-4o-mini", messages=[{"role": "user", "content": "hi"}], mock_response="ok")
+    assert seen == ["gpt-4o-mini"]
+    assert litellm.input_callback == []