[CI] Add mteb testing to test the accuracy of the embedding model (#17175)

Author: noooop

requirements/test.in

@@ -33,6 +33,7 @@ num2words # required for smolvlm test
 opencv-python-headless >= 4.11.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]==0.4.8 # required for model evaluation test
+mteb>=1.38.11, <2 # required for mteb test
 transformers==4.51.3
 tokenizers==0.21.1
 huggingface-hub[hf_xet]>=0.30.0  # Required for Xet downloads.

requirements/test.txt

@@ -99,6 +99,7 @@ datasets==3.0.2
     # via
     #   evaluate
     #   lm-eval
+    #   mteb
 decorator==5.1.1
     # via librosa
 dill==0.3.8
@@ -124,6 +125,8 @@ email-validator==2.2.0
     # via pydantic
 encodec==0.1.1
     # via vocos
+eval-type-backport==0.2.2
+    # via mteb
 evaluate==0.4.3
     # via lm-eval
 fastparquet==2024.11.0
@@ -291,6 +294,8 @@ msgpack==1.1.0
     # via
     #   librosa
     #   ray
+mteb==1.38.11
+    # via -r requirements/test.in
 multidict==6.1.0
     # via
     #   aiohttp
@@ -331,6 +336,7 @@ numpy==1.26.4
     #   librosa
     #   matplotlib
     #   mistral-common
+    #   mteb
     #   numba
     #   numexpr
     #   opencv-python-headless
@@ -443,6 +449,8 @@ plotly==5.24.1
     # via genai-perf
 pluggy==1.5.0
     # via pytest
+polars==1.29.0
+    # via mteb
 pooch==1.8.2
     # via librosa
 portalocker==2.10.1
@@ -476,6 +484,7 @@ pydantic==2.9.2
     # via
     #   datamodel-code-generator
     #   mistral-common
+    #   mteb
 pydantic-core==2.23.4
     # via pydantic
 pygments==2.18.0
@@ -522,6 +531,8 @@ python-dateutil==2.9.0.post0
     #   typepy
 python-rapidjson==1.20
     # via tritonclient
+pytrec-eval-terrier==0.5.7
+    # via mteb
 pytz==2024.2
     # via
     #   pandas
@@ -564,6 +575,7 @@ requests==2.32.3
     #   huggingface-hub
     #   lm-eval
     #   mistral-common
+    #   mteb
     #   pooch
     #   ray
     #   responses
@@ -580,6 +592,7 @@ rfc3987==1.3.8
 rich==13.9.4
     # via
     #   genai-perf
+    #   mteb
     #   typer
 rouge-score==0.1.2
     # via lm-eval
@@ -607,16 +620,20 @@ scikit-learn==1.5.2
     # via
     #   librosa
     #   lm-eval
+    #   mteb
     #   sentence-transformers
 scipy==1.13.1
     # via
     #   librosa
+    #   mteb
     #   scikit-learn
     #   sentence-transformers
     #   statsmodels
     #   vocos
 sentence-transformers==3.2.1
-    # via -r requirements/test.in
+    # via
+    #   -r requirements/test.in
+    #   mteb
 sentencepiece==0.2.0
     # via mistral-common
 setuptools==77.0.3
@@ -696,6 +713,7 @@ torch==2.7.0+cu128
     #   fastsafetensors
     #   lm-eval
     #   mamba-ssm
+    #   mteb
     #   peft
     #   runai-model-streamer
     #   sentence-transformers
@@ -720,6 +738,7 @@ tqdm==4.66.6
     #   evaluate
     #   huggingface-hub
     #   lm-eval
+    #   mteb
     #   nltk
     #   peft
     #   pqdm
@@ -759,6 +778,7 @@ typing-extensions==4.12.2
     #   huggingface-hub
     #   librosa
     #   mistral-common
+    #   mteb
     #   pqdm
     #   pydantic
     #   pydantic-core

tests/entrypoints/openai/correctness/test_mteb.py

@@ -0,0 +1,42 @@
+# SPDX-License-Identifier: Apache-2.0
+import math
+import os
+
+import pytest
+
+from tests.models.language.pooling.mteb_utils import (MTEB_EMBED_TASKS,
+                                                      OpenAIClientMtebEncoder,
+                                                      run_mteb_embed_task,
+                                                      run_mteb_embed_task_st)
+from tests.utils import RemoteOpenAIServer
+
+os.environ["VLLM_LOGGING_LEVEL"] = "WARNING"
+
+MODEL_NAME = "BAAI/bge-m3"
+DTYPE = "float16"
+MAIN_SCORE = 0.7873427091972599
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--task", "embed", "--dtype", DTYPE, "--enforce-eager",
+        "--max-model-len", "512"
+    ]
+
+    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+        yield remote_server
+
+
+def test_mteb(server):
+    client = server.get_client()
+    encoder = OpenAIClientMtebEncoder(MODEL_NAME, client)
+    vllm_main_score = run_mteb_embed_task(encoder, MTEB_EMBED_TASKS)
+    st_main_score = MAIN_SCORE or run_mteb_embed_task_st(
+        MODEL_NAME, MTEB_EMBED_TASKS)
+
+    print("VLLM main score: ", vllm_main_score)
+    print("SentenceTransformer main score: ", st_main_score)
+    print("Difference: ", st_main_score - vllm_main_score)
+
+    assert math.isclose(st_main_score, vllm_main_score, rel_tol=1e-4)

tests/models/language/pooling/test_gte.py

@@ -58,8 +58,6 @@
 @pytest.mark.parametrize("model_info", MODELS)
 def test_models_mteb(hf_runner, vllm_runner,
                      model_info: EmbedModelInfo) -> None:
-    pytest.skip("Skipping mteb test.")
-
     from .mteb_utils import mteb_test_embed_models
 
     vllm_extra_kwargs: dict[str, Any] = {}

tests/models/language/pooling/test_nomic.py

@@ -23,7 +23,6 @@
 @pytest.mark.parametrize("model_info", MODELS)
 def test_models_mteb(hf_runner, vllm_runner,
                      model_info: EmbedModelInfo) -> None:
-    pytest.skip("Skipping mteb test.")
     from .mteb_utils import mteb_test_embed_models
     mteb_test_embed_models(hf_runner, vllm_runner, model_info)
 

tests/models/language/pooling/test_snowflake_arctic_embed.py

@@ -46,7 +46,6 @@ def test_models_mteb(
     vllm_runner,
     model_info: EmbedModelInfo,
 ) -> None:
-    pytest.skip("Skipping mteb test.")
     from .mteb_utils import mteb_test_embed_models
     mteb_test_embed_models(hf_runner, vllm_runner, model_info)