[CI/Build] Split up models tests (vllm-project#10069)

Signed-off-by: DarkLight1337 <[email protected]>

.buildkite/test-pipeline.yaml

@@ -305,7 +305,7 @@ steps:
 
 #####  models test  #####
 
-- label: Basic Models Test # 3min
+- label: Basic Models Test # 10min
   source_file_dependencies:
   - vllm/
   - tests/models
@@ -314,23 +314,24 @@ steps:
     - 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 (Standard) # 35min
+- label: Decoder-only Language Models Test (Standard) # 18min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/language
   commands:
-    - pytest -v -s models/decoder_only/language/test_models.py
+    - pytest -v -s models/decoder_only/language -m core_model
+    - pytest -v -s models/decoder_only/language -m quant_model
 
-- label: Decoder-only Language Models Test (Extended) # 1h20min
+- label: Decoder-only Language Models Test (Extended) # 46min
   nightly: true
   source_file_dependencies:
   - vllm/
   - tests/models/decoder_only/language
   commands:
-    - pytest -v -s models/decoder_only/language --ignore=models/decoder_only/language/test_models.py
+    - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
 
-- label: Decoder-only Multi-Modal Models Test (Standard) # 26min
+- label: Decoder-only Multi-Modal Models Test (Standard) # 22min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
@@ -339,21 +340,24 @@ steps:
   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
 
-- label: Decoder-only Multi-Modal Models Test (Extended)
+- label: Decoder-only Multi-Modal Models Test (Extended) # 1h10m
   nightly: true
   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 -m 'not core_model'
+    - 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'
+    - 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 # 6min
+- label: Other Models Test # 20min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/

pyproject.toml

@@ -95,6 +95,7 @@ markers = [
     "skip_global_cleanup",
     "core_model: enable this model test in each PR instead of only nightly",
     "cpu_model: enable this model test in CPU tests",
+    "quant_model: run this model test under Quantized category",
     "distributed_2_gpus: run this test only in distributed tests for 2 GPUs",
     "skip_v1: do not run this test with v1",
 ]

tests/models/decoder_only/language/test_aqlm.py

@@ -38,6 +38,7 @@
 ]
 
 
+@pytest.mark.quant_model
 @pytest.mark.skipif(not is_quant_method_supported("aqlm"),
                     reason="AQLM is not supported on this GPU type.")
 @pytest.mark.parametrize("model", ["ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"])

tests/models/decoder_only/language/test_fp8.py

@@ -15,6 +15,7 @@
 os.environ["TOKENIZERS_PARALLELISM"] = "true"
 
 
+@pytest.mark.quant_model
 @pytest.mark.skipif(not is_quant_method_supported("fp8"),
                     reason="fp8 is not supported on this GPU type.")
 @pytest.mark.parametrize(

tests/models/decoder_only/language/test_gguf.py

@@ -17,26 +17,21 @@
 
 MAX_MODEL_LEN = 1024
 
-# FIXME: Move this to confest
-MODELS = [
-    ("meta-llama/Llama-3.2-1B-Instruct",
-     hf_hub_download("bartowski/Llama-3.2-1B-Instruct-GGUF",
-                     filename="Llama-3.2-1B-Instruct-Q4_K_M.gguf")),
-    ("meta-llama/Llama-3.2-1B-Instruct",
-     hf_hub_download("bartowski/Llama-3.2-1B-Instruct-GGUF",
-                     filename="Llama-3.2-1B-Instruct-IQ4_XS.gguf")),
-    ("Qwen/Qwen2-1.5B-Instruct",
-     hf_hub_download("Qwen/Qwen2-1.5B-Instruct-GGUF",
-                     filename="qwen2-1_5b-instruct-q4_k_m.gguf")),
-    ("Qwen/Qwen2-1.5B-Instruct",
-     hf_hub_download("legraphista/Qwen2-1.5B-Instruct-IMat-GGUF",
-                     filename="Qwen2-1.5B-Instruct.IQ4_XS.gguf")),
-]
-
 
 @pytest.mark.skipif(not is_quant_method_supported("gguf"),
                     reason="gguf is not supported on this GPU type.")
-@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize(("original_model", "gguf_id", "gguf_path"), [
+    ("meta-llama/Llama-3.2-1B-Instruct",
+     "bartowski/Llama-3.2-1B-Instruct-GGUF",
+     "Llama-3.2-1B-Instruct-Q4_K_M.gguf"),
+    ("meta-llama/Llama-3.2-1B-Instruct",
+     "bartowski/Llama-3.2-1B-Instruct-GGUF",
+     "Llama-3.2-1B-Instruct-IQ4_XS.gguf"),
+    ("Qwen/Qwen2-1.5B-Instruct", "Qwen/Qwen2-1.5B-Instruct-GGUF",
+     "qwen2-1_5b-instruct-q4_k_m.gguf"),
+    ("Qwen/Qwen2-1.5B-Instruct", "legraphista/Qwen2-1.5B-Instruct-IMat-GGUF",
+     "Qwen2-1.5B-Instruct.IQ4_XS.gguf"),
+])
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [32])
 @pytest.mark.parametrize("num_logprobs", [5])
@@ -45,7 +40,9 @@ def test_models(
     num_gpus_available,
     vllm_runner,
     example_prompts,
-    model,
+    original_model,
+    gguf_id,
+    gguf_path,
     dtype: str,
     max_tokens: int,
     num_logprobs: int,
@@ -54,7 +51,7 @@ def test_models(
     if num_gpus_available < tp_size:
         pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
 
-    original_model, gguf_model = model
+    gguf_model = hf_hub_download(gguf_id, filename=gguf_path)
 
     tokenizer = AutoTokenizer.from_pretrained(original_model)
     messages = [[{

tests/models/decoder_only/language/test_gptq_marlin.py

@@ -33,6 +33,7 @@
 ]
 
 
+@pytest.mark.quant_model
 @pytest.mark.flaky(reruns=3)
 @pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
                     reason="gptq_marlin is not supported on this GPU type.")

tests/models/decoder_only/language/test_gptq_marlin_24.py

@@ -38,6 +38,7 @@ class ModelPair:
 ]
 
 
+@pytest.mark.quant_model
 @pytest.mark.flaky(reruns=2)
 @pytest.mark.skipif(not is_quant_method_supported("gptq_marlin_24"),
                     reason="Marlin24 is not supported on this GPU type.")

tests/models/decoder_only/language/test_granite.py

@@ -7,7 +7,9 @@
 from ...utils import check_logprobs_close
 
 MODELS = [
+    # TODO(sang): Sliding window should be tested separately.
     "ibm/PowerLM-3b",
+    "ibm/PowerMoE-3b",
 ]
 
 
@@ -24,7 +26,6 @@ def test_models(
     max_tokens: int,
     num_logprobs: int,
 ) -> None:
-    # TODO(sang): Sliding window should be tested separately.
     with hf_runner(model, dtype=dtype) as hf_model:
         hf_outputs = hf_model.generate_greedy_logprobs_limit(
             example_prompts, max_tokens, num_logprobs)

tests/models/decoder_only/language/test_modelopt.py

@@ -39,6 +39,7 @@
 @pytest.mark.skip(
     reason=
     "Prevent unstable test based on golden strings from breaking the build.")
+@pytest.mark.quant_model
 @pytest.mark.skipif(not is_quant_method_supported("fp8"),
                     reason="fp8 is not supported on this GPU type.")
 @pytest.mark.parametrize("model_name", MODELS)

tests/models/decoder_only/language/test_models.py

@@ -1,8 +1,5 @@
 """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_models.py`.
 """
 import pytest
@@ -35,6 +32,7 @@
 target_dtype = "half"
 
 
+@pytest.mark.core_model
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", [target_dtype])
 @pytest.mark.parametrize("max_tokens", [32])

tests/models/decoder_only/vision_language/mm_processor_kwargs/test_llava_next.py

@@ -56,11 +56,13 @@ def test_dummy_data_for_llava_next_feature_size(dummy_data_for_llava_next,
     ctx.model_config.hf_config.image_grid_pinpoints = gridpoints
     seq_len = 5000  # bigger than the max feature size for any image
 
-    seq_data, mm_data = dummy_data_for_llava_next(
+    dummy_data = dummy_data_for_llava_next(
         ctx,
         seq_len=seq_len,
         mm_counts={"image": 1},
     )
+    seq_data = dummy_data.seq_data
+    mm_data = dummy_data.multi_modal_data
 
     # The dummy data dims should match the gridpoint with the biggest feat size
     assert mm_data["image"].height == expected_size[0]

tests/models/decoder_only/vision_language/mm_processor_kwargs/test_phi3v.py

@@ -131,12 +131,13 @@ def test_dummy_data_override(dummy_data_for_phi3v, model: str, num_crops: int,
         mm_processor_kwargs=None,
     )
 
-    sequence_data, _, = dummy_data_for_phi3v(
+    dummy_data = dummy_data_for_phi3v(
         ctx=ctx,
         seq_len=8192,  # Should be bigger than num_imgs * toks_per_img
         mm_counts={"image": num_imgs},
         num_crops=num_crops,
     )
+    sequence_data = dummy_data.seq_data
     # Ensure we have the right number of placeholders per num_crops size
     img_tok_count = sequence_data.get_token_ids().count(_IMAGE_TOKEN_ID)
     assert img_tok_count == toks_per_img * num_imgs

tests/models/decoder_only/vision_language/mm_processor_kwargs/test_qwen2_vl.py

@@ -86,10 +86,17 @@ def test_qwen2_vl_dummy_data(dummy_data_for_qwen2_vl,
 
     # NOTE: video value is required, but isn't actually used
     # when making the dummy data except for error handling currently
-    seq_data, mm_data = dummy_data_for_qwen2_vl(qwen2_vl_context, seq_len, {
-        "image": 1,
-        "video": 0
-    }, **mm_processor_kwargs)
+    dummy_data = dummy_data_for_qwen2_vl(
+        ctx=qwen2_vl_context,
+        seq_len=seq_len,
+        mm_counts={
+            "image": 1,
+            "video": 0
+        },
+        **mm_processor_kwargs,
+    )
+    seq_data = dummy_data.seq_data
+    mm_data = dummy_data.multi_modal_data
 
     # Ensure we have the right number of placeholders for min/max pixel values
     assert seq_data.get_token_ids().count(image_token_id) == token_count

tests/models/decoder_only/vision_language/test_internvl.py → tests/models/decoder_only/vision_language/test_awq.py

@@ -1,4 +1,4 @@
-from typing import List, Optional, Tuple, Type
+from typing import List, Optional, Type
 
 import pytest
 import torch
@@ -19,7 +19,8 @@
 def run_awq_test(
     vllm_runner: Type[VllmRunner],
     image_assets: _ImageAssets,
-    models: Tuple[str, str],
+    source_model: str,
+    quant_model: str,
     *,
     size_factors: List[float],
     dtype: str,
@@ -28,8 +29,6 @@ def run_awq_test(
     tensor_parallel_size: int,
     distributed_executor_backend: Optional[str] = None,
 ):
-    source_model, quant_model = models
-
     images = [asset.pil_image for asset in image_assets]
 
     inputs_per_image = [(
@@ -84,8 +83,11 @@ def run_awq_test(
         )
 
 
+@pytest.mark.quant_model
 @pytest.mark.parametrize(
-    "models", [("OpenGVLab/InternVL2-2B", "OpenGVLab/InternVL2-2B-AWQ")])
+    ("source_model", "quant_model"),
+    [("OpenGVLab/InternVL2-2B", "OpenGVLab/InternVL2-2B-AWQ")],
+)
 @pytest.mark.parametrize(
     "size_factors",
     [
@@ -103,12 +105,13 @@ def run_awq_test(
 @pytest.mark.parametrize("max_tokens", [128])
 @pytest.mark.parametrize("num_logprobs", [5])
 @torch.inference_mode()
-def test_awq_models(vllm_runner, image_assets, models, size_factors,
-                    dtype: str, max_tokens: int, num_logprobs: int) -> None:
+def test_awq_models(vllm_runner, image_assets, source_model, quant_model,
+                    size_factors, dtype, max_tokens, num_logprobs) -> None:
     run_awq_test(
         vllm_runner,
         image_assets,
-        models,
+        source_model,
+        quant_model,
         size_factors=size_factors,
         dtype=dtype,
         max_tokens=max_tokens,