From 13710577330649bb1263336f1a60948302ae8c76 Mon Sep 17 00:00:00 2001 From: Afloat Date: Wed, 11 Nov 2026 21:56:23 +0800 Subject: [PATCH] fix(llm): preserve native Ollama token usage --- src/llm/providers/ollama.py | 14 ++++++ tests/test_ollama_usage.py | 92 +++++++++++++++++++++++++++++++++++++ 2 files changed, 106 insertions(+) create mode 100644 tests/test_ollama_usage.py diff --git a/src/llm/providers/ollama.py b/src/llm/providers/ollama.py index 0a7434daa..28aefa4c9 100644 --- a/src/llm/providers/ollama.py +++ b/src/llm/providers/ollama.py @@ -31,6 +31,19 @@ from llm.providers.reasoning import prefilled_reasoning_from_env, strip_reasonin logger = logging.getLogger(__name__) +def _token_usage(response: dict[str, Any]) -> dict[str, int] | None: + """Expose measured counters using the OpenAI-compatible usage field names.""" + fields = (("prompt_tokens", "prompt_eval_count"), ("completion_tokens", "eval_count")) + usage = { + name: value + for name, source in fields + if type(value := response.get(source)) is int and value >= 0 + } + if len(usage) == len(fields): + return {**usage, "total_tokens": usage["prompt_tokens"] + usage["completion_tokens"]} + return usage or None + + def _messages_payload(messages: list[Message]) -> list[dict[str, Any]]: """Serialize messages, naming the tool each tool result answers. @@ -190,6 +203,7 @@ class OllamaProvider(LLMProvider): content=content, tool_calls=tool_calls, model=model, + usage=_token_usage(result), stop_reason=result.get("done_reason"), ) except Exception as e: diff --git a/tests/test_ollama_usage.py b/tests/test_ollama_usage.py new file mode 100644 index 000000000..f7e7c2a81 --- /dev/null +++ b/tests/test_ollama_usage.py @@ -0,0 +1,92 @@ +from __future__ import annotations + +import json +import urllib.request +from io import BytesIO +from typing import Any + +import pytest + +from llm.core.types import LLMInput, Message, Role +from llm.providers.ollama import OllamaProvider + +pytestmark = pytest.mark.unit + + +@pytest.mark.parametrize( + "counters, expected", + [ + ( + {"prompt_eval_count": 142, "eval_count": 57}, + {"prompt_tokens": 142, "completion_tokens": 57, "total_tokens": 199}, + ), + ( + {"prompt_eval_count": 0, "eval_count": 0}, + {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}, + ), + ({"prompt_eval_count": 7}, {"prompt_tokens": 7}), + ({"eval_count": 5}, {"completion_tokens": 5}), + ({}, None), + ({"prompt_eval_count": True, "eval_count": -1}, None), + ({"prompt_eval_count": "12", "eval_count": 4}, {"completion_tokens": 4}), + ({"prompt_eval_count": 1.5, "eval_count": None}, None), + ], +) +def test_native_ollama_counters_reach_output_usage( + monkeypatch: pytest.MonkeyPatch, + counters: dict[str, Any], + expected: dict[str, int] | None, +) -> None: + response = { + "model": "llama3.2", + "message": {"role": "assistant", "content": "result"}, + "done": True, + "done_reason": "stop", + **counters, + } + + def native_response(request: urllib.request.Request, timeout: float) -> BytesIO: + assert request.full_url == "http://localhost:11434/api/chat" + assert request.get_method() == "POST" + assert isinstance(request.data, bytes) + assert json.loads(request.data)["stream"] is False + assert timeout == 60 + return BytesIO(json.dumps(response).encode("utf-8")) + + monkeypatch.setattr(urllib.request, "urlopen", native_response) + output = OllamaProvider(base_url="http://localhost:11434", default_model="llama3.2").generate( + LLMInput(messages=[Message(role=Role.USER, content="hello")]) + ) + + assert output.usage == expected + assert output.to_dict().get("usage") == expected + assert output.content == "result" + assert output.model == "llama3.2" + assert output.stop_reason == "stop" + + +def test_tool_call_response_retains_usage(monkeypatch: pytest.MonkeyPatch) -> None: + response = { + "message": { + "role": "assistant", + "content": "", + "tool_calls": [{"function": {"name": "lookup", "arguments": {"key": "value"}}}], + }, + "done_reason": "stop", + "prompt_eval_count": 10, + "eval_count": 3, + } + def native_response(request: urllib.request.Request, timeout: float) -> BytesIO: + return BytesIO(json.dumps(response).encode("utf-8")) + + monkeypatch.setattr(urllib.request, "urlopen", native_response) + + output = OllamaProvider(default_model="llama3.2").generate( + LLMInput(messages=[Message(role=Role.USER, content="lookup")]) + ) + + assert output.usage == {"prompt_tokens": 10, "completion_tokens": 3, "total_tokens": 13} + assert output.has_tool_calls + assert output.tool_calls is not None + assert output.tool_calls[0].name == "lookup" + assert output.tool_calls[0].arguments == {"key": "value"}