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

Author: DarkLight1337

.buildkite/run-cpu-test-ppc64le.sh

@@ -27,9 +27,9 @@ function cpu_tests() {
       decord einops librosa peft Pillow sentence-transformers soundfile \
       transformers_stream_generator matplotlib datamodel_code_generator
     pip install torchvision --index-url https://download.pytorch.org/whl/cpu
-    pytest -v -s tests/models/embedding/language
-    pytest -v -s tests/models/encoder_decoder/language
-    pytest -v -s tests/models/decoder_only/language/test_models.py
+    pytest -v -s tests/models/decoder_only/language -m cpu_model
+    pytest -v -s tests/models/embedding/language -m cpu_model
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
     pytest -v -s tests/models/decoder_only/audio_language -m cpu_model
     pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"
 

.buildkite/run-cpu-test.sh

@@ -38,9 +38,9 @@ function cpu_tests() {
       decord einops librosa peft Pillow sentence-transformers soundfile \
       transformers_stream_generator matplotlib datamodel_code_generator
     pip install torchvision --index-url https://download.pytorch.org/whl/cpu
-    pytest -v -s tests/models/embedding/language
-    pytest -v -s tests/models/encoder_decoder/language
-    pytest -v -s tests/models/decoder_only/language/test_models.py
+    pytest -v -s tests/models/decoder_only/language -m cpu_model
+    pytest -v -s tests/models/embedding/language -m cpu_model
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
     pytest -v -s tests/models/decoder_only/audio_language -m cpu_model
     pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"
 

.buildkite/test-pipeline.yaml

@@ -323,62 +323,60 @@ steps:
     - pytest -v -s models/test_registry.py
     - pytest -v -s models/test_initialization.py
 
-- label: Decoder-only Language Models Test (Standard) # 18min
+- label: Language Models Test (Standard) # 42min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/language
+  - tests/models/embedding/language
+  - tests/models/encoder_decoder/language
   commands:
-    - pytest -v -s models/decoder_only/language -m core_model
-    - pytest -v -s models/decoder_only/language -m quant_model
+    - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
+    - pytest -v -s models/embedding/language -m core_model
+    - pytest -v -s models/embedding/vision_language -m core_model
 
-- label: Decoder-only Language Models Test (Extended) # 46min
+- label: Language Models Test (Extended) # 50min
   nightly: true
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/language
+  - tests/models/embedding/language
+  - tests/models/encoder_decoder/language
   commands:
     - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
+    - pytest -v -s models/embedding/language -m 'not core_model'
+    - pytest -v -s models/embedding/vision_language -m 'not core_model'
 
-- label: Decoder-only Multi-Modal Models Test (Standard) # 22min
+- label: Multi-Modal Models Test (Standard) # 26min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/audio_language
   - tests/models/decoder_only/vision_language
+  - tests/models/embedding/vision_language
+  - tests/models/encoder_decoder/vision_language
   commands:
-    - pytest -v -s models/decoder_only/audio_language -m core_model
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m core_model
-    # No tests under this group for now
-    # - pytest -v -s models/decoder_only/audio_language -m quant_model
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m quant_model
+    - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
+    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
+    - pytest -v -s models/encoder_decoder/language -m core_model
+    - pytest -v -s models/encoder_decoder/vision_language -m core_model
 
-- label: Decoder-only Multi-Modal Models Test (Extended) # 1h10m
+- label: Multi-Modal Models Test (Extended) # 1h15m
   nightly: true
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/audio_language
   - tests/models/decoder_only/vision_language
+  - tests/models/embedding/vision_language
+  - tests/models/encoder_decoder/vision_language
   commands:
     - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
     # HACK - run phi3v tests separately to sidestep this transformers bug
     # https://github.com/huggingface/transformers/issues/34307
     - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
     - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
-
-- label: Other Models Test # 20min
-  #mirror_hardwares: [amd]
-  source_file_dependencies:
-  - vllm/
-  - tests/models/embedding/language
-  - tests/models/embedding/vision_language
-  - tests/models/encoder_decoder/language
-  - tests/models/encoder_decoder/vision_language
-  commands:
-    - pytest -v -s models/embedding/language
-    - pytest -v -s models/embedding/vision_language
-    - pytest -v -s models/encoder_decoder/language
-    - pytest -v -s models/encoder_decoder/vision_language
+    - pytest -v -s models/encoder_decoder/language -m 'not core_model'
+    - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test

docs/source/models/supported_models.rst

@@ -330,11 +330,16 @@ Text Embedding
     - :code:`BAAI/bge-multilingual-gemma2`, etc.
     - 
     - ✅︎
-  * - :code:`MistralModel`
-    - Mistral-based
+  * - :code:`LlamaModel`, :code:`LlamaForCausalLM`, :code:`MistralModel`, etc.
+    - Llama-based
     - :code:`intfloat/e5-mistral-7b-instruct`, etc.
     - ✅︎
     - ✅︎
+  * - :code:`Qwen2Model`, :code:`Qwen2ForCausalLM`
+    - Qwen2-based
+    - :code:`ssmits/Qwen2-7B-Instruct-embed-base`, :code:`Alibaba-NLP/gte-Qwen2-1.5B-instruct`, etc.
+    - ✅︎
+    - ✅︎
 
 .. important::
   Some model architectures support both generation and embedding tasks.
@@ -355,7 +360,7 @@ Reward Modeling
   * - :code:`Qwen2ForRewardModel`
     - Qwen2-based
     - :code:`Qwen/Qwen2.5-Math-RM-72B`, etc.
-    - 
+    - ✅︎
     - ✅︎
 
 .. note::
@@ -376,7 +381,7 @@ Classification
   * - :code:`Qwen2ForSequenceClassification`
     - Qwen2-based
     - :code:`jason9693/Qwen2.5-1.5B-apeach`, etc.
-    - 
+    - ✅︎
     - ✅︎
 
 .. note::

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 @@
 """
 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(

tests/models/embedding/language/test_cls_models.py

@@ -9,10 +9,14 @@
 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)