vllm/tests/models/embedding/utils.py

from typing import List, Sequence

import torch
import torch.nn.functional as F


def check_embeddings_close(
    *,
    embeddings_0_lst: Sequence[List[float]],
    embeddings_1_lst: Sequence[List[float]],
    name_0: str,
    name_1: str,
    tol: float = 1e-3,
) -> None:
    assert len(embeddings_0_lst) == len(embeddings_1_lst)

    for prompt_idx, (embeddings_0, embeddings_1) in enumerate(
            zip(embeddings_0_lst, embeddings_1_lst)):
        assert len(embeddings_0) == len(embeddings_1), (
            f"Length mismatch: {len(embeddings_0)} vs. {len(embeddings_1)}")

        sim = F.cosine_similarity(torch.tensor(embeddings_0),
                                  torch.tensor(embeddings_1),
                                  dim=0)

        fail_msg = (f"Test{prompt_idx}:"
                    f"\n{name_0}:\t{embeddings_0!r}"
                    f"\n{name_1}:\t{embeddings_1!r}")

        assert sim >= 1 - tol, fail_msg