Files
ECC/tests/test_ollama_usage.py
T

93 lines
3.1 KiB
Python

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"}