9fb900f90c
[CI/Test] improve robustness of test by replacing del with context manager (hf_runner) (#5347)
44 lines
1.2 KiB
Python
44 lines
1.2 KiB
Python
"""Compare the outputs of HF and vLLM for Mistral models using greedy sampling.
|
|
|
|
Run `pytest tests/models/test_llama_embedding.py`.
|
|
"""
|
|
import pytest
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
MODELS = [
|
|
"intfloat/e5-mistral-7b-instruct",
|
|
]
|
|
|
|
|
|
def compare_embeddings(embeddings1, embeddings2):
|
|
similarities = [
|
|
F.cosine_similarity(torch.tensor(e1), torch.tensor(e2), dim=0)
|
|
for e1, e2 in zip(embeddings1, embeddings2)
|
|
]
|
|
return similarities
|
|
|
|
|
|
@pytest.mark.parametrize("model", MODELS)
|
|
@pytest.mark.parametrize("dtype", ["half"])
|
|
def test_models(
|
|
hf_runner,
|
|
vllm_runner,
|
|
example_prompts,
|
|
model: str,
|
|
dtype: str,
|
|
) -> None:
|
|
with hf_runner(model, dtype=dtype, is_embedding_model=True) as hf_model:
|
|
hf_outputs = hf_model.encode(example_prompts)
|
|
|
|
vllm_model = vllm_runner(model, dtype=dtype)
|
|
vllm_outputs = vllm_model.encode(example_prompts)
|
|
del vllm_model
|
|
|
|
similarities = compare_embeddings(hf_outputs, vllm_outputs)
|
|
all_similarities = torch.stack(similarities)
|
|
tolerance = 1e-2
|
|
assert torch.all((all_similarities <= 1.0 + tolerance)
|
|
& (all_similarities >= 1.0 - tolerance)
|
|
), f"Not all values are within {tolerance} of 1.0"
|