1234567891011121314151617181920212223242526272829303132333435363738394041 |
- """Compare the outputs of HF and vLLM for Mistral models using greedy sampling.
- Run `pytest tests/models/test_mistral.py`.
- """
- import pytest
- from .utils import check_logprobs_close
- MODELS = [
- "mistralai/Mistral-7B-Instruct-v0.1",
- "mistralai/Mistral-7B-Instruct-v0.3",
- ]
- @pytest.mark.parametrize("model", MODELS)
- @pytest.mark.parametrize("dtype", ["bfloat16"])
- @pytest.mark.parametrize("max_tokens", [64])
- @pytest.mark.parametrize("num_logprobs", [5])
- def test_models(
- hf_runner,
- aphrodite_runner,
- example_prompts,
- model: str,
- dtype: str,
- max_tokens: int,
- num_logprobs: int,
- ) -> None:
- # TODO: Sliding window should be tested separately.
- with hf_runner(model, dtype=dtype) as hf_model:
- hf_outputs = hf_model.generate_greedy_logprobs_limit(
- example_prompts, max_tokens, num_logprobs)
- with aphrodite_runner(model, dtype=dtype) as aphrodite_model:
- aphrodite_outputs = aphrodite_model.generate_greedy_logprobs(
- example_prompts, max_tokens, num_logprobs)
- check_logprobs_close(
- outputs_0_lst=hf_outputs,
- outputs_1_lst=aphrodite_outputs,
- name_0="hf",
- name_1="aphrodite",
- )
|