[CI/Build] Add test decorator for minimum GPU memory (#8925)

This commit is contained in:
Cyrus Leung
2024-09-29 02:50:51 +00:00
committed by GitHub
parent d081da0064
commit 26a68d5d7e
14 changed files with 117 additions and 73 deletions
@@ -7,6 +7,7 @@ import torch
from vllm.utils import is_cpu
from ....utils import large_gpu_test
from ...utils import check_logprobs_close
MODELS = [
@@ -69,20 +70,10 @@ def test_phimoe_routing_function():
assert torch.equal(topk_ids, ground_truth[test_id]["topk_ids"])
def get_gpu_memory():
try:
props = torch.cuda.get_device_properties(torch.cuda.current_device())
gpu_memory = props.total_memory / (1024**3)
return gpu_memory
except Exception:
return 0
@pytest.mark.skipif(condition=is_cpu(),
reason="This test takes a lot time to run on CPU, "
"and vllm CI's disk space is not enough for this model.")
@pytest.mark.skipif(condition=get_gpu_memory() < 100,
reason="Skip this test if GPU memory is insufficient.")
@large_gpu_test(min_gb=80)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [64])
@@ -11,6 +11,7 @@ from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
from ....conftest import (VIDEO_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_VideoAssets)
from ....utils import large_gpu_test
from ...utils import check_logprobs_close
# Video test
@@ -164,9 +165,7 @@ def run_video_test(
)
@pytest.mark.skip(
reason=
"Model is too big, test passed on L40 locally but will OOM on CI machine.")
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"size_factors",
@@ -210,9 +209,7 @@ def test_models(hf_runner, vllm_runner, video_assets, model, size_factors,
)
@pytest.mark.skip(
reason=
"Model is too big, test passed on L40 locally but will OOM on CI machine.")
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"sizes",
@@ -306,9 +303,7 @@ def run_image_test(
)
@pytest.mark.skip(
reason=
"Model is too big, test passed on L40 locally but will OOM on CI machine.")
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [128])
@@ -17,7 +17,7 @@ from vllm import EngineArgs, LLMEngine, SamplingParams, TokensPrompt
from vllm.multimodal import MultiModalDataBuiltins
from vllm.sequence import Logprob, SampleLogprobs
from ....utils import VLLM_PATH
from ....utils import VLLM_PATH, large_gpu_test
from ...utils import check_logprobs_close
if TYPE_CHECKING:
@@ -121,10 +121,7 @@ def load_outputs_w_logprobs(filename: "StrPath") -> OutputsLogprobs:
for tokens, text, logprobs in json_data]
@pytest.mark.skip(
reason=
"Model is too big, test passed on A100 locally but will OOM on CI machine."
)
@large_gpu_test(min_gb=80)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("max_model_len", MAX_MODEL_LEN)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@@ -157,10 +154,7 @@ def test_chat(
name_1="output")
@pytest.mark.skip(
reason=
"Model is too big, test passed on A100 locally but will OOM on CI machine."
)
@large_gpu_test(min_gb=80)
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
def test_model_engine(vllm_runner, model: str, dtype: str) -> None:
@@ -9,6 +9,7 @@ from vllm.sequence import SampleLogprobs
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from ....utils import large_gpu_test
from ...utils import check_logprobs_close
_LIMIT_IMAGE_PER_PROMPT = 1
@@ -227,29 +228,26 @@ def _run_test(
)
SIZES = [
# Text only
[],
# Single-size
[(512, 512)],
# Single-size, batched
[(512, 512), (512, 512), (512, 512)],
# Multi-size, batched
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028)],
# Multi-size, batched, including text only
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028), None],
# mllama has 8 possible aspect ratios, carefully set the sizes
# to cover all of them
]
@pytest.mark.skip(
reason=
"Model is too big, test passed on L40 locally but will OOM on CI machine.")
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("sizes", SIZES)
@pytest.mark.parametrize(
"sizes",
[
# Text only
[],
# Single-size
[(512, 512)],
# Single-size, batched
[(512, 512), (512, 512), (512, 512)],
# Multi-size, batched
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028)],
# Multi-size, batched, including text only
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028), None],
# mllama has 8 possible aspect ratios, carefully set the sizes
# to cover all of them
])
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])