| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Sequence |
| 4 | |
| 5 | import openai |
| 6 | import pytest |
| 7 | |
| 8 | from tests.conftest import HfRunner |
| 9 | from tests.models.utils import EmbedModelInfo, check_embeddings_close, matryoshka_fy |
| 10 | from vllm.entrypoints.pooling.embed.protocol import EmbeddingResponse |
| 11 | |
| 12 | |
| 13 | def float_embeddings(response: EmbeddingResponse) -> list[list[float]]: |
| 14 | embeddings = [] |
| 15 | for d in response.data: |
| 16 | assert isinstance(d.embedding, list) |
| 17 | embeddings.append(d.embedding) |
| 18 | return embeddings |
| 19 | |
| 20 | |
| 21 | def run_embedding_correctness_test( |
| 22 | hf_model: "HfRunner", |
| 23 | inputs: list[str], |
| 24 | vllm_outputs: Sequence[list[float]], |
| 25 | dimensions: int | None = None, |
| 26 | ): |
| 27 | hf_outputs = hf_model.encode(inputs) |
| 28 | if dimensions: |
| 29 | hf_outputs = matryoshka_fy(hf_outputs, dimensions) |
| 30 | |
| 31 | check_embeddings_close( |
| 32 | embeddings_0_lst=hf_outputs, |
| 33 | embeddings_1_lst=vllm_outputs, |
| 34 | name_0="hf", |
| 35 | name_1="vllm", |
| 36 | tol=1e-2, |
| 37 | ) |
| 38 | |
| 39 | |
| 40 | def correctness_test_embed_models( |
| 41 | hf_runner, |
| 42 | vllm_runner, |
| 43 | model_info: EmbedModelInfo, |
| 44 | example_prompts, |
| 45 | vllm_extra_kwargs=None, |
| 46 | hf_model_callback=None, |
| 47 | ): |
| 48 | pytest.skip("Debug only, ci prefers to use mteb test.") |
| 49 | |
| 50 | # The example_prompts has ending "\n", for example: |
| 51 | # "Write a short story about a robot that dreams for the first time.\n" |
| 52 | # sentence_transformers will strip the input texts, see: |
| 53 | # https://github.com/UKPLab/sentence-transformers/blob/v3.1.1/sentence_transformers/models/Transformer.py#L159 |
| 54 | # This makes the input_ids different between hf_model and vllm_model. |
| 55 | # So we need to strip the input texts to avoid test failing. |
| 56 | example_prompts = [str(s).strip() for s in example_prompts] |
| 57 | |
| 58 | vllm_extra_kwargs = vllm_extra_kwargs or {} |
| 59 | vllm_extra_kwargs["dtype"] = model_info.dtype |
| 60 | |
| 61 | if model_info.hf_overrides is not None: |
| 62 | vllm_extra_kwargs["hf_overrides"] = model_info.hf_overrides |
| 63 | |
| 64 | with vllm_runner( |
| 65 | model_info.name, runner="pooling", max_model_len=None, **vllm_extra_kwargs |
| 66 | ) as vllm_model: |
| 67 | vllm_outputs = vllm_model.embed(example_prompts) |
| 68 | |
| 69 | with hf_runner( |
| 70 | model_info.name, |
| 71 | dtype=model_info.hf_dtype, |
| 72 | is_sentence_transformer=True, |
| 73 | ) as hf_model: |
| 74 | if hf_model_callback is not None: |
| 75 | hf_model_callback(hf_model) |
| 76 | |
| 77 | run_embedding_correctness_test(hf_model, example_prompts, vllm_outputs) |
| 78 | |
| 79 | |
| 80 | async def run_client_embeddings( |
| 81 | client: openai.AsyncOpenAI, |
| 82 | model_name: str, |
| 83 | queries: list[str], |
| 84 | instruction: str = "", |
| 85 | ) -> list[list[float]]: |
| 86 | outputs = await client.embeddings.create( |
| 87 | model=model_name, |
| 88 | input=[instruction + q for q in queries], |
| 89 | ) |
| 90 | return [data.embedding for data in outputs.data] |