[CI/Build] Reorganize models tests

.buildkite/run-cpu-test.sh

@@ -23,12 +23,10 @@ docker exec cpu-test-avx2 bash -c "python3 examples/offline_inference.py"
 # Run basic model test
 docker exec cpu-test bash -c "
   pip install pytest matplotlib einops transformers_stream_generator
-  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py \
-      --ignore=tests/models/test_oot_registration.py \
-      --ignore=tests/models/test_registry.py \
-      --ignore=tests/models/test_fp8.py \
-      --ignore=tests/models/test_jamba.py \
-      --ignore=tests/models/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
+  pytest -v -s tests/models/decoder_only/language \
+    --ignore=tests/models/test_fp8.py \
+    --ignore=tests/models/decoder_only/language/test_jamba.py \
+    --ignore=tests/models/decoder_only/language/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
 
 # Run compressed-tensor test
 docker exec cpu-test bash -c "

.buildkite/test-pipeline.yaml

@@ -94,7 +94,6 @@ steps:
   - pytest -v -s entrypoints/test_chat_utils.py
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
-
 - label: Distributed Tests (4 GPUs) # 10min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -164,30 +163,13 @@ steps:
     - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
     - python3 offline_inference_encoder_decoder.py
 
-- label: Models Test # 1hr10min
-  source_file_dependencies:
-  - vllm/
-  - tests/models
-  commands:
-    - pip install -e ./plugins/vllm_add_dummy_model
-    - pytest -v -s models/test_oot_registration.py # it needs a clean process
-    - pytest -v -s models -m \"not vlm\" --ignore=models/test_oot_registration.py
-
 - label: torch compile integration test
   source_file_dependencies:
   - vllm/
   commands:
     - pytest -v -s ./compile/test_full_graph.py
     - pytest -v -s ./compile/test_wrapper.py
 
-
-- label: Vision Language Models Test # 42min
-  #mirror_hardwares: [amd]
-  source_file_dependencies:
-  - vllm/
-  commands:
-    - pytest -v -s models -m vlm
-
 - label: Prefix Caching Test # 7min
   #mirror_hardwares: [amd]
   source_file_dependencies:
@@ -286,6 +268,45 @@ steps:
   commands:
     - pytest -v -s tool_use
 
+#####  models test  #####
+
+- label: Basic Models Test # 3min
+  source_file_dependencies:
+  - vllm/
+  - tests/models
+  commands:
+    - pip install -e ./plugins/vllm_add_dummy_model
+    - pytest -v -s models/test_oot_registration.py # it needs a clean process
+    - pytest -v -s models/*.py --ignore=models/test_oot_registration.py
+
+- label: Decoder-only Language Models Test # 1h3min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/models/decoder_only/language
+  commands:
+    - pytest -v -s models/decoder_only/language
+
+- label: Decoder-only Multi-Modal Models Test # 56min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/models/decoder_only/audio_language
+  - tests/models/decoder_only/vision_language
+  commands:
+    - pytest -v -s models/decoder_only/audio_language
+    - pytest -v -s models/decoder_only/vision_language
+
+- label: Other Models Test # 5min
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/
+  - tests/models/embedding/language
+  - tests/models/encoder_decoder/language
+  commands:
+    - pytest -v -s models/embedding/language
+    - pytest -v -s models/encoder_decoder/language
+
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -311,11 +332,11 @@ steps:
   - tests/distributed/
   commands:
   - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
   - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
 
 - label: Distributed Tests (2 GPUs) # 28min
   #mirror_hardwares: [amd]
@@ -328,11 +349,10 @@ steps:
   - vllm/model_executor/models/
   - tests/distributed/
   commands:
-  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
-  - TARGET_TEST_SUITE=L4 pytest -v -s distributed/test_basic_distributed_correctness.py
-  - pytest -v -s distributed/test_basic_distributed_correctness_enc_dec.py
-  - pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - pytest -v -s distributed/test_multimodal_broadcast.py
+  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep -q 'Same node test passed'
+  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m distributed_2_gpus
+  # Avoid importing model tests that cause CUDA reinitialization error
+  - pytest models/encoder_decoder/language/test_bart.py models/decoder_only/vision_language/test_broadcast.py -v -s -m distributed_2_gpus
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py

docs/source/models/supported_models.rst

@@ -342,7 +342,7 @@ Note that, as an inference engine, vLLM does not introduce new models. Therefore
 
 We have the following levels of testing for models:
 
-1. **Strict Consistency**: We compare the output of the model with the output of the model in the HuggingFace Transformers library under greedy decoding. This is the most stringent test. Please refer to `test_models.py <https://github.com/vllm-project/vllm/blob/main/tests/models/test_models.py>`_ and `test_big_models.py <https://github.com/vllm-project/vllm/blob/main/tests/models/test_big_models.py>`_ for the models that have passed this test.
+1. **Strict Consistency**: We compare the output of the model with the output of the model in the HuggingFace Transformers library under greedy decoding. This is the most stringent test. Please refer to `models tests <https://github.com/vllm-project/vllm/blob/main/tests/models>`_ for the models that have passed this test.
 2. **Output Sensibility**: We check if the output of the model is sensible and coherent, by measuring the perplexity of the output and checking for any obvious errors. This is a less stringent test.
 3. **Runtime Functionality**: We check if the model can be loaded and run without errors. This is the least stringent test. Please refer to `functionality tests <https://github.com/vllm-project/vllm/tree/main/tests>`_ and `examples <https://github.com/vllm-project/vllm/tree/main/examples>`_ for the models that have passed this test.
 4. **Community Feedback**: We rely on the community to provide feedback on the models. If a model is broken or not working as expected, we encourage users to raise issues to report it or open pull requests to fix it. The rest of the models fall under this category.

pyproject.toml

@@ -85,5 +85,6 @@ skip_gitignore = true
 [tool.pytest.ini_options]
 markers = [
     "skip_global_cleanup",
-    "vlm: run tests for vision language models only",
+    "core_model: run this model test in each PR instead of just daily",
+    "distributed_2_gpus: run this test only in distributed tests for 2 GPUs",
 ]

tests/basic_correctness/test_basic_correctness.py

@@ -15,12 +15,15 @@
 from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
 
 from ..models.utils import check_outputs_equal
+from ..utils import multi_gpu_test
 
 MODELS = [
     "facebook/opt-125m",
     "meta-llama/Llama-2-7b-hf",
 ]
 
+TARGET_TEST_SUITE = os.environ.get("TARGET_TEST_SUITE", "L4")
+
 
 def test_vllm_gc_ed():
     """Verify vllm instance is GC'ed when it is deleted"""
@@ -70,6 +73,65 @@ def test_models(
     )
 
 
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize(
+    "model, distributed_executor_backend, attention_backend, "
+    "test_suite", [
+        ("facebook/opt-125m", "ray", "", "L4"),
+        ("facebook/opt-125m", "mp", "", "L4"),
+        ("meta-llama/Llama-2-7b-hf", "ray", "", "L4"),
+        ("meta-llama/Llama-2-7b-hf", "mp", "", "L4"),
+        ("facebook/opt-125m", "ray", "", "A100"),
+        ("facebook/opt-125m", "mp", "", "A100"),
+        ("facebook/opt-125m", "mp", "FLASHINFER", "A100"),
+        ("meta-llama/Meta-Llama-3-8B", "ray", "FLASHINFER", "A100"),
+    ])
+def test_models_distributed(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    distributed_executor_backend: str,
+    attention_backend: str,
+    test_suite: str,
+) -> None:
+
+    if test_suite != TARGET_TEST_SUITE:
+        pytest.skip(f"Skip test for {test_suite}")
+
+    if model == "meta-llama/Llama-2-7b-hf" and distributed_executor_backend == "ray" and attention_backend == "" and test_suite == "L4":  # noqa
+        # test ray adag
+        os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
+        os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
+
+    if attention_backend:
+        os.environ["VLLM_ATTENTION_BACKEND"] = attention_backend
+
+    dtype = "half"
+    max_tokens = 5
+
+    # NOTE: take care of the order. run vLLM first, and then run HF.
+    # vLLM needs a fresh new process without cuda initialization.
+    # if we run HF first, the cuda initialization will be done and it
+    # will hurt multiprocessing backend with fork method (the default method).
+    with vllm_runner(model,
+                     dtype=dtype,
+                     tensor_parallel_size=2,
+                     distributed_executor_backend=distributed_executor_backend
+                     ) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
 def test_model_with_failure(vllm_runner) -> None:
     try:
         with patch("vllm.model_executor.models.opt.OPTForCausalLM.forward",

tests/basic_correctness/test_chunked_prefill.py

@@ -6,11 +6,13 @@
 
 Run `pytest tests/models/test_chunked_prefill.py`.
 """
+import os
 from contextlib import nullcontext
 
 import pytest
 
 from ..models.utils import check_logprobs_close, check_outputs_equal
+from ..utils import multi_gpu_test
 
 MODELS = [
     "facebook/opt-125m",
@@ -66,6 +68,59 @@ def test_models(
     )
 
 
+@multi_gpu_test(num_gpus=2)
+@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
+@pytest.mark.parametrize("model", MODELS)
+def test_models_distributed(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    model: str,
+    distributed_executor_backend: str,
+) -> None:
+    if (model == "meta-llama/Llama-2-7b-hf"
+            and distributed_executor_backend == "ray"):
+        # test ray adag
+        os.environ['VLLM_USE_RAY_SPMD_WORKER'] = "1"
+        os.environ['VLLM_USE_RAY_COMPILED_DAG'] = "1"
+
+    dtype = "half"
+    max_tokens = 5
+    chunked_prefill_token_size = 16
+
+    # Add a chunked prefill config.
+    max_num_seqs = min(chunked_prefill_token_size, 256)
+    assert chunked_prefill_token_size != -1
+    enable_chunked_prefill = True
+    max_num_batched_tokens = chunked_prefill_token_size
+
+    # NOTE: take care of the order. run vLLM first, and then run HF.
+    # vLLM needs a fresh new process without cuda initialization.
+    # if we run HF first, the cuda initialization will be done and it
+    # will hurt multiprocessing backend with fork method (the default method).
+
+    with vllm_runner(
+            model,
+            dtype=dtype,
+            tensor_parallel_size=2,
+            max_num_seqs=max_num_seqs,
+            enable_chunked_prefill=enable_chunked_prefill,
+            max_num_batched_tokens=max_num_batched_tokens,
+            distributed_executor_backend=distributed_executor_backend,
+    ) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
+
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs = hf_model.generate_greedy(example_prompts, max_tokens)
+
+    check_outputs_equal(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )
+
+
 @pytest.mark.parametrize(
     "kv_cache_dtype,model",
     [("fp8_e4m3",

tests/basic_correctness/test_preemption.py

@@ -19,10 +19,13 @@
     "facebook/opt-125m",
 ]
 
-assert ENABLE_ARTIFICIAL_PREEMPT is True, (
-    "Use an env var VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1. "
-    "`VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest "
-    "tests/basic_correctness/test_preemption.py`")
+
+@pytest.fixture(scope="module", autouse=True)
+def check_settings():
+    assert ENABLE_ARTIFICIAL_PREEMPT is True, (
+        "Use an env var VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1. "
+        "`VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest "
+        "tests/basic_correctness/test_preemption.py`")
 
 
 @pytest.fixture

tests/conftest.py

@@ -6,8 +6,8 @@
 import tempfile
 from collections import UserList
 from enum import Enum
-from typing import (Any, Callable, Dict, List, Optional, Tuple, TypedDict,
-                    TypeVar, Union)
+from typing import (Any, Callable, Dict, List, Optional, Tuple, Type,
+                    TypedDict, TypeVar, Union)
 
 import numpy as np
 import pytest
@@ -18,6 +18,7 @@
 from PIL import Image
 from transformers import (AutoModelForCausalLM, AutoTokenizer, BatchEncoding,
                           BatchFeature)
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from vllm import LLM, SamplingParams
 from vllm.assets.image import ImageAsset
@@ -260,7 +261,7 @@ def __init__(
         *,
         model_kwargs: Optional[Dict[str, Any]] = None,
         is_embedding_model: bool = False,
-        auto_cls=AutoModelForCausalLM,
+        auto_cls: Type[_BaseAutoModelClass] = AutoModelForCausalLM,
         postprocess_inputs: Callable[[BatchEncoding],
                                      BatchEncoding] = identity,
     ) -> None:
@@ -292,20 +293,14 @@ def __init__(
             trust_remote_code=True,
         )
 
-        try:
-            # don't put this import at the top level
-            # it will call torch.cuda.device_count()
-            from transformers import AutoProcessor  # noqa: F401
-            self.processor = AutoProcessor.from_pretrained(
-                model_name,
-                torch_dtype=torch_dtype,
-                trust_remote_code=True,
-            )
-        except Exception as exc:
-            logger.warning(
-                "Unable to auto-load HuggingFace processor for model (%s). "
-                "Using tokenizer instead. Reason: %s", model_name, exc)
-            self.processor = self.tokenizer
+        # don't put this import at the top level
+        # it will call torch.cuda.device_count()
+        from transformers import AutoProcessor  # noqa: F401
+        self.processor = AutoProcessor.from_pretrained(
+            model_name,
+            torch_dtype=torch_dtype,
+            trust_remote_code=True,
+        )
 
         self.postprocess_inputs = postprocess_inputs