diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index c5fcbfb2564..5971ed25b7f 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -15664,9 +15664,24 @@ async def _count_request_tokens( ) system_message: Final = _system_message(system) prompt_turn: Final = prompt_only and not (system_message is None and not tools) - typed_messages: Final = cast( # cast-ok: request messages are raw chat-shaped dicts that token_counter normalizes + from litellm.google_genai.adapters.transformation import GoogleGenAIAdapter + + typed_messages: Final = cast( # cast-ok: request messages and adapter output are chat-shaped dicts Sequence[AllMessageValues] | None, - [{"role": "user", "content": prompt}] if prompt_turn else (None if prompt_only else request_messages), + [{"role": "user", "content": prompt}] + if prompt_turn + else ( + None + if prompt_only + else request_messages + if request_messages is not None or contents is None + else GoogleGenAIAdapter() + .translate_generate_content_to_completion( # pyright: ignore[reportUnknownMemberType] # the adapter method has an untyped **kwargs + model_to_use, + cast(list[dict[str, object]], contents), # cast-ok: the request contents are Gemini content dicts + ) + .get("messages") + ), ) counted_messages: Final = ( typed_messages if typed_messages is None or system_message is None else (system_message, *typed_messages) diff --git a/tests/unit/proxy/test_proxy_token_counter.py b/tests/unit/proxy/test_proxy_token_counter.py index 68eb05e60dd..aa9f01b6046 100644 --- a/tests/unit/proxy/test_proxy_token_counter.py +++ b/tests/unit/proxy/test_proxy_token_counter.py @@ -4252,3 +4252,55 @@ def test_utils_token_counter_line_stays_one_line_when_the_model_carries_a_newlin assert line["request_model"] == forged_model assert "\n" not in message assert message.startswith(f'token count - litellm_call_id="count log" request_model={json.dumps(forged_model)} ') + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "contents,messages", + [ + ( + [{"parts": [{"text": "Count every word in this request."}]}], + [{"role": "user", "content": "Count every word in this request."}], + ), + ( + [{"role": "user", "parts": [{"text": "One part."}, {"text": "Another part."}]}], + [ + { + "role": "user", + "content": [{"type": "text", "text": "One part."}, {"type": "text", "text": "Another part."}], + } + ], + ), + ( + [{"role": "user", "parts": [{"text": "Hello"}]}, {"role": "model", "parts": [{"text": "Welcome back"}]}], + [{"role": "user", "content": "Hello"}, {"role": "assistant", "content": "Welcome back"}], + ), + ], +) +async def test_token_counter_counts_gemini_contents( + contents: list[dict[str, object]], messages: list[dict[str, object]] +) -> None: + expected: Final = await token_counter( + TokenCountRequest(model="gemini/gemini-3.8-flash", messages=messages) + ) + actual: Final = await token_counter( + TokenCountRequest(model="gemini/gemini-3.8-flash", contents=contents) + ) + assert actual.total_tokens == expected.total_tokens + assert actual.total_tokens > 0 + + +@pytest.mark.asyncio +async def test_token_counter_prefers_explicit_messages_over_contents() -> None: + messages: Final = [{"role": "user", "content": "Count this message."}] + expected: Final = await token_counter( + TokenCountRequest(model="gemini/gemini-3.8-flash", messages=messages) + ) + actual: Final = await token_counter( + TokenCountRequest( + model="gemini/gemini-3.8-flash", + messages=messages, + contents=[{"parts": [{"text": "Do not count this different text."}]}], + ) + ) + assert actual.total_tokens == expected.total_tokens