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

tests/models/decoder_only/language/test_jamba.py

@@ -33,6 +33,10 @@ def test_models(
 
     with vllm_runner(model, dtype=dtype) as vllm_model:
         vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+        # 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)
 
     for i in range(len(example_prompts)):
         hf_output_ids, hf_output_str = hf_outputs[i]
@@ -293,17 +297,3 @@ def test_jamba_distributed_produces_identical_generation(
         name_0="vllm_tp_1",
         name_1="vllm_tp_2",
     )
-
-
-@pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("dtype", ["float"])
-def test_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/decoder_only/language/test_mamba.py

@@ -51,6 +51,10 @@ def test_models(
 
     with vllm_runner(model, dtype=dtype) as vllm_model:
         vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+        # 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)
 
     for i in range(len(example_prompts)):
         hf_output_ids, hf_output_str = hf_outputs[i]
@@ -279,17 +283,3 @@ def test_state_cleanup(
     except ValueError:
         pytest.fail("Mamba inner state wasn't cleaned up between states, "
                     "could be related to finished_requests_ids")
-
-
-@pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("dtype", ["float"])
-def test_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/decoder_only/language/test_models.py

@@ -4,37 +4,52 @@ Run `pytest tests/models/test_models.py`.
 """
 import pytest
 
-from vllm.platforms import current_platform
-
 from ...utils import check_logprobs_close
 
-MODELS = [
-    "facebook/opt-125m",  # opt
-    "openai-community/gpt2",  # gpt2
-    # "Milos/slovak-gpt-j-405M",  # gptj
-    # "bigcode/tiny_starcoder_py",  # gpt_bigcode
-    # "EleutherAI/pythia-70m",  # gpt_neox
-    "bigscience/bloom-560m",  # bloom - testing alibi slopes
-    "microsoft/phi-2",  # phi
-    # "stabilityai/stablelm-3b-4e1t",  # stablelm
-    # "bigcode/starcoder2-3b",  # starcoder2
-    "google/gemma-1.1-2b-it",  # gemma
-    "Qwen/Qwen2.5-0.5B-Instruct",  # qwen2
-    "meta-llama/Llama-3.2-1B-Instruct",  # llama
-]
 
-if not current_platform.is_cpu():
-    MODELS += [
-        # fused_moe which not supported on CPU
-        "openbmb/MiniCPM3-4B",
-    ]
-
-target_dtype = "half"
-
-
-@pytest.mark.core_model
-@pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("dtype", [target_dtype])
+@pytest.mark.parametrize(
+    "model",
+    [
+        pytest.param(
+            "bigscience/bloom-560m",  # bloom - testing alibi slopes
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
+        ),
+        pytest.param(
+            "openai-community/gpt2",  # gpt2
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
+        ),
+        pytest.param("Milos/slovak-gpt-j-405M"),  # gptj
+        pytest.param("bigcode/tiny_starcoder_py"),  # gpt_bigcode
+        pytest.param("EleutherAI/pythia-70m"),  # gpt_neox
+        pytest.param(
+            "google/gemma-1.1-2b-it",  # gemma
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
+        ),
+        pytest.param(
+            "meta-llama/Llama-3.2-1B-Instruct",  # llama
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
+        ),
+        pytest.param(
+            "openbmb/MiniCPM3-4B",
+            # fused_moe not supported on CPU
+            marks=[pytest.mark.core_model],
+        ),
+        pytest.param(
+            "facebook/opt-125m",  # opt
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
+        ),
+        pytest.param(
+            "microsoft/phi-2",  # phi
+            marks=[pytest.mark.core_model],
+        ),
+        pytest.param(
+            "Qwen/Qwen2.5-0.5B-Instruct",  # qwen2
+            marks=[pytest.mark.core_model],
+        ),
+        pytest.param("stabilityai/stablelm-3b-4e1t"),  # stablelm
+        pytest.param("bigcode/starcoder2-3b"),  # starcoder2
+    ])
+@pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [32])
 @pytest.mark.parametrize("num_logprobs", [5])
 def test_models(