[CI] Add mteb testing for rerank models (#19344)

Author: noooop

requirements/test.in

@@ -33,7 +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
+mteb[bm25s]>=1.38.11, <2 # required for mteb test
 transformers==4.52.4
 tokenizers==0.21.1
 huggingface-hub[hf_xet]>=0.30.0  # Required for Xet downloads.

requirements/test.txt

@@ -51,6 +51,8 @@ black==24.10.0
     # via datamodel-code-generator
 blobfile==3.0.0
     # via -r requirements/test.in
+bm25s==0.2.13
+    # via mteb
 boto3==1.35.57
     # via tensorizer
 botocore==1.35.57
@@ -344,6 +346,7 @@ numpy==1.26.4
     #   -r requirements/test.in
     #   accelerate
     #   bitsandbytes
+    #   bm25s
     #   contourpy
     #   cupy-cuda12x
     #   datasets
@@ -534,6 +537,8 @@ pyparsing==3.2.0
     # via matplotlib
 pyrate-limiter==3.7.0
     # via schemathesis
+pystemmer==3.0.0
+    # via mteb
 pytablewriter==1.2.0
     # via lm-eval
 pytest==8.3.3
@@ -668,6 +673,7 @@ scikit-learn==1.5.2
     #   sentence-transformers
 scipy==1.13.1
     # via
+    #   bm25s
     #   librosa
     #   mteb
     #   scikit-learn

tests/conftest.py

@@ -727,8 +727,12 @@ def encode(self, prompts: list[str], *args,
                **kwargs) -> list[list[torch.Tensor]]:
         return self.model.encode(prompts, *args, **kwargs)
 
-    def predict(self, prompts: list[list[str]]) -> torch.Tensor:
-        return self.model.predict(prompts, convert_to_tensor=True)
+    def predict(self, prompts: list[list[str]], *args,
+                **kwargs) -> torch.Tensor:
+        return self.model.predict(prompts,
+                                  *args,
+                                  convert_to_tensor=True,
+                                  **kwargs)
 
     def __enter__(self):
         return self
@@ -1037,8 +1041,10 @@ def score(
         self,
         text_1: Union[str, list[str]],
         text_2: Union[str, list[str]],
+        *args,
+        **kwargs,
     ) -> list[float]:
-        req_outputs = self.model.score(text_1, text_2)
+        req_outputs = self.model.score(text_1, text_2, *args, **kwargs)
         return [req_output.outputs.score for req_output in req_outputs]
 
     def apply_model(self, func: Callable[[nn.Module], _R]) -> list[_R]:

…rypoints/openai/correctness/test_mteb.py …ts/openai/correctness/test_mteb_embed.pytests/entrypoints/openai/correctness/test_mteb.py renamed to tests/entrypoints/openai/correctness/test_mteb_embed.py tests/entrypoints/openai/correctness/test_mteb.py renamed to tests/entrypoints/openai/correctness/test_mteb_embed.py

@@ -7,34 +7,30 @@
 from tests.models.language.pooling.mteb_utils import (MTEB_EMBED_TASKS,
                                                       MTEB_EMBED_TOL,
                                                       OpenAIClientMtebEncoder,
-                                                      run_mteb_embed_task,
-                                                      run_mteb_embed_task_st)
+                                                      run_mteb_embed_task)
 from tests.utils import RemoteOpenAIServer
 
 os.environ["VLLM_LOGGING_LEVEL"] = "WARNING"
 
-MODEL_NAME = "BAAI/bge-m3"
-DTYPE = "float16"
-MAIN_SCORE = 0.7873427091972599
+MODEL_NAME = "intfloat/e5-small"
+MAIN_SCORE = 0.7422994752439667
 
 
 @pytest.fixture(scope="module")
 def server():
     args = [
-        "--task", "embed", "--dtype", DTYPE, "--enforce-eager",
-        "--max-model-len", "512"
+        "--task", "embed", "--enforce-eager", "--disable-uvicorn-access-log"
     ]
 
     with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
         yield remote_server
 
 
-def test_mteb(server):
+def test_mteb_embed(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)
+    st_main_score = MAIN_SCORE
 
     print("VLLM main score: ", vllm_main_score)
     print("SentenceTransformer main score: ", st_main_score)

tests/entrypoints/openai/correctness/test_mteb_score.py

@@ -0,0 +1,59 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import os
+
+import pytest
+
+# yapf conflicts with isort for this block
+# yapf: disable
+from tests.models.language.pooling.mteb_utils import (MTEB_RERANK_LANGS,
+                                                      MTEB_RERANK_TASKS,
+                                                      MTEB_RERANK_TOL,
+                                                      RerankClientMtebEncoder,
+                                                      ScoreClientMtebEncoder,
+                                                      run_mteb_rerank)
+# yapf: enable
+from tests.utils import RemoteOpenAIServer
+
+os.environ["VLLM_LOGGING_LEVEL"] = "WARNING"
+
+MODEL_NAME = "cross-encoder/ms-marco-MiniLM-L-6-v2"
+MAIN_SCORE = 0.33437
+
+
+@pytest.fixture(scope="module")
+def server():
+    args = [
+        "--task", "score", "--enforce-eager", "--disable-uvicorn-access-log"
+    ]
+
+    with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
+        yield remote_server
+
+
+def test_mteb_score(server):
+    url = server.url_for("score")
+    encoder = ScoreClientMtebEncoder(MODEL_NAME, url)
+    vllm_main_score = run_mteb_rerank(encoder, MTEB_RERANK_TASKS,
+                                      MTEB_RERANK_LANGS)
+    st_main_score = MAIN_SCORE
+
+    print("VLLM main score: ", vllm_main_score)
+    print("SentenceTransformer main score: ", st_main_score)
+    print("Difference: ", st_main_score - vllm_main_score)
+
+    assert st_main_score == pytest.approx(vllm_main_score, abs=MTEB_RERANK_TOL)
+
+
+def test_mteb_rerank(server):
+    url = server.url_for("rerank")
+    encoder = RerankClientMtebEncoder(MODEL_NAME, url)
+    vllm_main_score = run_mteb_rerank(encoder, MTEB_RERANK_TASKS,
+                                      MTEB_RERANK_LANGS)
+    st_main_score = MAIN_SCORE
+
+    print("VLLM main score: ", vllm_main_score)
+    print("SentenceTransformer main score: ", st_main_score)
+    print("Difference: ", st_main_score - vllm_main_score)
+
+    assert st_main_score == pytest.approx(vllm_main_score, abs=MTEB_RERANK_TOL)