add qwen3

vllm-v0.6.2/tests/models/embedding/language/test_cls_models.py

@@ -0,0 +1,45 @@
+"""Compare the outputs of HF and vLLM when using greedy sampling.
+
+This test only tests small models. Big models such as 7B should be tested from
+test_big_models.py because it could use a larger instance to run tests.
+
+Run `pytest tests/models/test_cls_models.py`.
+"""
+import pytest
+import torch
+from transformers import AutoModelForSequenceClassification
+
+
+@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,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+) -> 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)
+
+    # 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)

vllm-v0.6.2/tests/models/embedding/language/test_embedding.py

@@ -0,0 +1,60 @@
+"""Compare the embedding outputs of HF and vLLM models.
+
+Run `pytest tests/models/embedding/language/test_embedding.py`.
+"""
+import pytest
+
+from ..utils import check_embeddings_close
+
+
+@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,
+    vllm_runner,
+    example_prompts,
+    model,
+    dtype: str,
+) -> None:
+    # 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:
+    # https://github.com/UKPLab/sentence-transformers/blob/v3.1.1/sentence_transformers/models/Transformer.py#L159
+    # This makes the input_ids different between hf_model and vllm_model.
+    # So we need to strip the input texts to avoid test failing.
+    example_prompts = [str(s).strip() for s in example_prompts]
+
+    with hf_runner(model, dtype=dtype,
+                   is_sentence_transformer=True) as hf_model:
+        hf_outputs = hf_model.encode(example_prompts)
+
+    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,
+        embeddings_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+        tol=1e-2,
+    )