[CI] improve embed testing (#18747)

tests/entrypoints/openai/correctness/test_mteb.py

@@ -4,6 +4,7 @@ import os
 import pytest
 
 from tests.models.language.pooling.mteb_utils import (MTEB_EMBED_TASKS,
+                                                      MTEB_EMBED_TOL,
                                                       OpenAIClientMtebEncoder,
                                                       run_mteb_embed_task,
                                                       run_mteb_embed_task_st)
@@ -38,4 +39,4 @@ def test_mteb(server):
     print("SentenceTransformer main score: ", st_main_score)
     print("Difference: ", st_main_score - vllm_main_score)
 
-    assert st_main_score == pytest.approx(vllm_main_score, rel=1e-4)
+    assert st_main_score == pytest.approx(vllm_main_score, abs=MTEB_EMBED_TOL)

tests/entrypoints/openai/test_embedding.py

@@ -11,7 +11,8 @@ import requests
 from vllm.entrypoints.openai.protocol import EmbeddingResponse
 from vllm.transformers_utils.tokenizer import get_tokenizer
 
-from ...models.utils import run_embedding_correctness_test
+from ...models.language.pooling.embed_utils import (
+    run_embedding_correctness_test)
 from ...utils import RemoteOpenAIServer
 
 MODEL_NAME = "intfloat/multilingual-e5-small"

tests/entrypoints/openai/test_embedding_dimensions.py

@@ -11,7 +11,9 @@ import pytest
 from vllm.entrypoints.openai.protocol import EmbeddingResponse
 
 from ...conftest import HfRunner
-from ...models.utils import EmbedModelInfo, run_embedding_correctness_test
+from ...models.language.pooling.embed_utils import (
+    run_embedding_correctness_test)
+from ...models.utils import EmbedModelInfo
 from ...utils import RemoteOpenAIServer
 
 MODELS = [