| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Run `pytest tests/entrypoints/openai/test_embedding_dimensions.py`.""" |
| 4 | |
| 5 | import openai |
| 6 | import pytest |
| 7 | |
| 8 | from tests.conftest import HfRunner |
| 9 | from tests.models.language.pooling.embed_utils import ( |
| 10 | float_embeddings, |
| 11 | run_embedding_correctness_test, |
| 12 | ) |
| 13 | from tests.models.utils import EmbedModelInfo |
| 14 | from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer |
| 15 | from vllm.entrypoints.pooling.embed.protocol import EmbeddingResponse |
| 16 | |
| 17 | MODELS = [ |
| 18 | EmbedModelInfo("intfloat/multilingual-e5-small", is_matryoshka=False), |
| 19 | EmbedModelInfo( |
| 20 | "Snowflake/snowflake-arctic-embed-m-v1.5", |
| 21 | is_matryoshka=True, |
| 22 | matryoshka_dimensions=[256], |
| 23 | ), |
| 24 | ] |
| 25 | |
| 26 | input_texts = [ |
| 27 | "The chef prepared a delicious meal.", |
| 28 | ] |
| 29 | |
| 30 | |
| 31 | @pytest.fixture(scope="module", params=MODELS) |
| 32 | def model_info(request): |
| 33 | return request.param |
| 34 | |
| 35 | |
| 36 | @pytest.fixture(scope="module", params=["bfloat16"]) |
| 37 | def dtype(request): |
| 38 | return request.param |
| 39 | |
| 40 | |
| 41 | @pytest.fixture(scope="module") |
| 42 | def server(model_info, dtype: str): |
| 43 | args = [ |
| 44 | "--runner", |
| 45 | "pooling", |
| 46 | # use half precision for speed and memory savings in CI environment |
| 47 | "--dtype", |
| 48 | dtype, |
| 49 | "--enforce-eager", |
| 50 | "--max-model-len", |
| 51 | "512", |
| 52 | ] + ROCM_EXTRA_ARGS |
| 53 | |
| 54 | if model_info.name == "Snowflake/snowflake-arctic-embed-m-v1.5": |
| 55 | # Manually enable Matryoshka Embeddings |
| 56 | args.extend( |
| 57 | ["--trust_remote_code", "--hf_overrides", '{"matryoshka_dimensions":[256]}'] |
| 58 | ) |
| 59 | |
| 60 | with RemoteOpenAIServer(model_info.name, args) as remote_server: |
| 61 | yield remote_server |
| 62 | |
| 63 | |
| 64 | @pytest.fixture(scope="module") |
| 65 | def hf_model(hf_runner, model_info, dtype: str): |
| 66 | with hf_runner( |
| 67 | model_info.name, dtype=dtype, is_sentence_transformer=True |
| 68 | ) as hf_model: |
| 69 | yield hf_model |
| 70 | |
| 71 | |
| 72 | @pytest.mark.asyncio |
| 73 | async def test_matryoshka( |
| 74 | model_info: EmbedModelInfo, server: RemoteOpenAIServer, hf_model: HfRunner |
| 75 | ): |
| 76 | client = server.get_async_client() |
| 77 | |
| 78 | async def make_request_and_correctness_test(dimensions): |
| 79 | prompts = input_texts * 3 |
| 80 | |
| 81 | embedding_response = await client.embeddings.create( |
| 82 | model=model_info.name, |
| 83 | input=prompts, |
| 84 | dimensions=dimensions, |
| 85 | encoding_format="float", |
| 86 | ) |
| 87 | embeddings = EmbeddingResponse.model_validate( |
| 88 | embedding_response.model_dump(mode="json") |
| 89 | ) |
| 90 | |
| 91 | assert embeddings.id is not None |
| 92 | assert len(embeddings.data) == 3 |
| 93 | assert len(embeddings.data[0].embedding) > 0 |
| 94 | assert embeddings.usage.completion_tokens == 0 |
| 95 | assert embeddings.usage.prompt_tokens > 0 |
| 96 | assert embeddings.usage.total_tokens > 0 |
| 97 | |
| 98 | if dimensions is not None: |
| 99 | assert len(embeddings.data[0].embedding) == dimensions |
| 100 | |
| 101 | vllm_outputs = float_embeddings(embeddings) |
| 102 | run_embedding_correctness_test(hf_model, prompts, vllm_outputs, dimensions) |
| 103 | |
| 104 | if model_info.is_matryoshka: |
| 105 | valid_dimensions: list[int | None] = [None] |
| 106 | if model_info.matryoshka_dimensions is not None: |
| 107 | valid_dimensions += model_info.matryoshka_dimensions[:2] |
| 108 | |
| 109 | for dimensions in valid_dimensions: |
| 110 | await make_request_and_correctness_test(dimensions) |
| 111 | |
| 112 | invalid_dimensions: list[int | None] = [-1] |
| 113 | if model_info.matryoshka_dimensions is not None: |
| 114 | assert 5 not in model_info.matryoshka_dimensions |
| 115 | invalid_dimensions.append(5) |
| 116 | |
| 117 | for dimensions in invalid_dimensions: |
| 118 | with pytest.raises(openai.BadRequestError): |
| 119 | await make_request_and_correctness_test(dimensions) |
| 120 | |
| 121 | else: |
| 122 | for dimensions in [None]: |
| 123 | await make_request_and_correctness_test(dimensions) |
| 124 | |
| 125 | for dimensions in [-1, 16]: |
| 126 | with pytest.raises(openai.BadRequestError): |
| 127 | await make_request_and_correctness_test(dimensions) |