[CI/Build] Update CPU tests to include all "standard" tests (#5481)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/models/decoder_only/audio_language/test_ultravox.py

@@ -5,11 +5,11 @@ import pytest
 import pytest_asyncio
 from transformers import AutoModel, AutoTokenizer, BatchEncoding
 
-from tests.utils import RemoteOpenAIServer
 from vllm.sequence import SampleLogprobs
 from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
 
 from ....conftest import HfRunner, VllmRunner
+from ....utils import RemoteOpenAIServer
 from ...utils import check_logprobs_close
 
 MODEL_NAME = "fixie-ai/ultravox-v0_3"
@@ -39,7 +39,10 @@ def audio(request):
     return AudioAsset(request.param)
 
 
-@pytest.fixture(params=({}, CHUNKED_PREFILL_KWARGS))
+@pytest.fixture(params=[
+    pytest.param({}, marks=pytest.mark.cpu_model),
+    pytest.param(CHUNKED_PREFILL_KWARGS),
+])
 def server(request, audio_assets):
     args = [
         "--dtype=bfloat16", "--max-model-len=4096", "--enforce-eager",
@@ -185,7 +188,10 @@ def run_multi_audio_test(
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [128])
 @pytest.mark.parametrize("num_logprobs", [5])
-@pytest.mark.parametrize("vllm_kwargs", [{}, CHUNKED_PREFILL_KWARGS])
+@pytest.mark.parametrize("vllm_kwargs", [
+    pytest.param({}, marks=pytest.mark.cpu_model),
+    pytest.param(CHUNKED_PREFILL_KWARGS),
+])
 def test_models(hf_runner, vllm_runner, audio, dtype: str, max_tokens: int,
                 num_logprobs: int, vllm_kwargs: dict) -> None:
 
@@ -207,7 +213,10 @@ def test_models(hf_runner, vllm_runner, audio, dtype: str, max_tokens: int,
 @pytest.mark.parametrize("dtype", ["half"])
 @pytest.mark.parametrize("max_tokens", [128])
 @pytest.mark.parametrize("num_logprobs", [5])
-@pytest.mark.parametrize("vllm_kwargs", [{}, CHUNKED_PREFILL_KWARGS])
+@pytest.mark.parametrize("vllm_kwargs", [
+    pytest.param({}, marks=pytest.mark.cpu_model),
+    pytest.param(CHUNKED_PREFILL_KWARGS),
+])
 def test_models_with_multiple_audios(vllm_runner, audio_assets, dtype: str,
                                      max_tokens: int, num_logprobs: int,
                                      vllm_kwargs: dict) -> None:

tests/models/decoder_only/vision_language/test_h2ovl.py

@@ -14,7 +14,6 @@ models = [
     "h2oai/h2ovl-mississippi-800m",  # Replace with your actual model names
     "h2oai/h2ovl-mississippi-2b",
 ]
-target_dtype = "bfloat16"
 
 
 def run_preprocessing_test(

tests/models/decoder_only/vision_language/test_models.py

@@ -94,7 +94,7 @@ VLM_TEST_SETTINGS = {
             ),
             limit_mm_per_prompt={"image": 4},
         )],
-        marks=[pytest.mark.core_model],
+        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
     ),
     "paligemma": VLMTestInfo(
         models=["google/paligemma-3b-mix-224"],
@@ -111,7 +111,8 @@ VLM_TEST_SETTINGS = {
             "pixel_values"
         ),
         vllm_output_post_proc=model_utils.paligemma_vllm_to_hf_output,
-        dtype="half" if current_platform.is_rocm() else ("half", "float"),
+        dtype=("half" if current_platform.is_cpu() or current_platform.is_rocm()
+               else ("half", "float")),
         marks=[pytest.mark.core_model],
     ),
     "qwen2_vl": VLMTestInfo(
@@ -128,7 +129,7 @@ VLM_TEST_SETTINGS = {
         max_num_seqs=2,
         auto_cls=AutoModelForVision2Seq,
         vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
-        marks=[pytest.mark.core_model],
+        marks=[pytest.mark.core_model, pytest.mark.cpu_model],
         image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
     ),
     #### Extended model tests
@@ -172,7 +173,6 @@ VLM_TEST_SETTINGS = {
         use_tokenizer_eos=True,
         vllm_output_post_proc=model_utils.fuyu_vllm_to_hf_output,
         num_logprobs=10,
-        dtype="bfloat16" if current_platform.is_cpu() else "half",
         image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
     ),
     "glm4": VLMTestInfo(
@@ -245,7 +245,6 @@ VLM_TEST_SETTINGS = {
         models=["llava-hf/llava-onevision-qwen2-0.5b-ov-hf"],
         test_type=VLMTestType.CUSTOM_INPUTS,
         prompt_formatter=lambda vid_prompt: f"<|im_start|>user\n{vid_prompt}<|im_end|>\n<|im_start|>assistant\n",   # noqa: E501
-        dtype="half",
         num_video_frames=16,
         max_model_len=16384,
         postprocess_inputs=model_utils.get_key_type_post_processor(
@@ -404,7 +403,6 @@ VLM_TEST_SETTINGS = {
         prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501
         test_type=VLMTestType.CUSTOM_INPUTS,
         max_model_len=4096,
-        dtype="bfloat16" if current_platform.is_cpu() else "half",
         use_tokenizer_eos=True,
         patch_hf_runner=model_utils.internvl_patch_hf_runner,
         custom_test_opts=[
@@ -419,7 +417,6 @@ VLM_TEST_SETTINGS = {
         test_type=VLMTestType.CUSTOM_INPUTS,
         max_model_len=16384,
         max_num_seqs=2,
-        dtype="half",
         postprocess_inputs=model_utils.get_key_type_post_processor(
             "pixel_values"
         ),

tests/models/decoder_only/vision_language/test_phi3v.py

@@ -44,8 +44,6 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
 
 
 target_dtype = "half"
-if current_platform.is_cpu():
-    target_dtype = "bfloat16"
 
 # ROCm Triton FA can run into shared memory issues with these models,
 # use other backends in the meantime