[Model] Support Qwen2 embeddings and use tags to select model tests (#10184)

tests/models/embedding/language/test_cls_models.py

@@ -9,10 +9,14 @@ import pytest
 import torch
 from transformers import AutoModelForSequenceClassification
 
-CLASSIFICATION_MODELS = ["jason9693/Qwen2.5-1.5B-apeach"]
 
-
-@pytest.mark.parametrize("model", CLASSIFICATION_MODELS)
+@pytest.mark.parametrize(
+    "model",
+    [
+        pytest.param("jason9693/Qwen2.5-1.5B-apeach",
+                     marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
+    ],
+)
 @pytest.mark.parametrize("dtype", ["float"])
 def test_classification_models(
     hf_runner,
@@ -23,31 +27,19 @@ def test_classification_models(
 ) -> None:
     with vllm_runner(model, dtype=dtype) as vllm_model:
         vllm_outputs = vllm_model.classify(example_prompts)
+        # This test is for verifying whether the model's extra_repr
+        # can be printed correctly.
+        print(vllm_model.model.llm_engine.model_executor.driver_worker.
+              model_runner.model)
 
     with hf_runner(model,
                    dtype=dtype,
                    auto_cls=AutoModelForSequenceClassification) as hf_model:
         hf_outputs = hf_model.classify(example_prompts)
 
-    print(hf_outputs, vllm_outputs)
-
     # check logits difference
     for hf_output, vllm_output in zip(hf_outputs, vllm_outputs):
         hf_output = torch.tensor(hf_output)
         vllm_output = torch.tensor(vllm_output)
 
         assert torch.allclose(hf_output, vllm_output, 1e-3)
-
-
-@pytest.mark.parametrize("model", CLASSIFICATION_MODELS)
-@pytest.mark.parametrize("dtype", ["float"])
-def test_classification_model_print(
-    vllm_runner,
-    model: str,
-    dtype: str,
-) -> None:
-    with vllm_runner(model, dtype=dtype) as vllm_model:
-        # This test is for verifying whether the model's extra_repr
-        # can be printed correctly.
-        print(vllm_model.model.llm_engine.model_executor.driver_worker.
-              model_runner.model)

tests/models/embedding/language/test_embedding.py

@@ -4,25 +4,25 @@ Run `pytest tests/models/embedding/language/test_embedding.py`.
 """
 import pytest
 
-from vllm.utils import current_platform
-
 from ..utils import check_embeddings_close
 
-# Model, Guard
-MODELS = [
-    "intfloat/e5-mistral-7b-instruct",
-    "BAAI/bge-base-en-v1.5",
-    "BAAI/bge-multilingual-gemma2",
-    "intfloat/multilingual-e5-large",
-]
 
-ENCODER_ONLY = [
-    "BAAI/bge-base-en-v1.5",
-    "intfloat/multilingual-e5-large",
-]
-
-
-@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize(
+    "model",
+    [
+        # [Encoder-only]
+        pytest.param("BAAI/bge-base-en-v1.5",
+                     marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
+        pytest.param("intfloat/multilingual-e5-large"),
+        # [Encoder-decoder]
+        pytest.param("intfloat/e5-mistral-7b-instruct",
+                     marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
+        pytest.param("BAAI/bge-multilingual-gemma2",
+                     marks=[pytest.mark.core_model]),
+        pytest.param("ssmits/Qwen2-7B-Instruct-embed-base"),
+        pytest.param("Alibaba-NLP/gte-Qwen2-1.5B-instruct"),
+    ],
+)
 @pytest.mark.parametrize("dtype", ["half"])
 def test_models(
     hf_runner,
@@ -31,9 +31,6 @@ def test_models(
     model,
     dtype: str,
 ) -> None:
-    if model not in ENCODER_ONLY and current_platform.is_cpu():
-        pytest.skip("Skip large embedding models test on CPU.")
-
     # The example_prompts has ending "\n", for example:
     # "Write a short story about a robot that dreams for the first time.\n"
     # sentence_transformers will strip the input texts, see:
@@ -46,8 +43,13 @@ def test_models(
                    is_sentence_transformer=True) as hf_model:
         hf_outputs = hf_model.encode(example_prompts)
 
-    with vllm_runner(model, dtype=dtype, max_model_len=None) as vllm_model:
+    with vllm_runner(model, task="embedding", dtype=dtype,
+                     max_model_len=None) as vllm_model:
         vllm_outputs = vllm_model.encode(example_prompts)
+        # This test is for verifying whether the model's extra_repr
+        # can be printed correctly.
+        print(vllm_model.model.llm_engine.model_executor.driver_worker.
+              model_runner.model)
 
     check_embeddings_close(
         embeddings_0_lst=hf_outputs,