biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build] Update models tests & examples (#8874)

Browse Source 
Co-authored-by: Roger Wang <ywang@roblox.com>
This commit is contained in:
Cyrus Leung
2024-09-29 00:54:35 +08:00
committed by GitHub
parent 19d02ff938
commit e1a3f5e831
13 changed files with 239 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files

153
tests/models/encoder_decoder/vision_language/test_mllama.py
View File

@@ -9,7 +9,6 @@ from vllm.sequence import SampleLogprobs
from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
_ImageAssets)
from ....utils import multi_gpu_test
from ...utils import check_logprobs_close
_LIMIT_IMAGE_PER_PROMPT = 1
@@ -47,14 +46,46 @@ def vllm_to_hf_output(vllm_output: Tuple[List[int], str,
if token_id != image_token_id or output_ids[idx - 1] != image_token_id
]
assert output_str[0] == " "
hf_output_str = output_str[1:]
hf_output_str = output_str
if hf_output_ids[-1] == eos_token_id:
hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
return hf_output_ids, hf_output_str, out_logprobs
def _get_inputs(
image_assets: _ImageAssets,
*,
size_factors: Optional[List[float]] = None,
sizes: Optional[List[Tuple[int, int]]] = None,
) -> List[Tuple[List[str], PromptImageInput]]:
images = [asset.pil_image for asset in image_assets]
if size_factors is not None:
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
elif sizes is not None:
inputs_per_image = [(
[
prompt if size is not None else text_only_prompts[0]
for size in sizes
],
[
image.resize(size) if size is not None else None
for size in sizes
],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
if len(sizes) == 0:
inputs_per_image.append(
(text_only_prompts, [None] * len(text_only_prompts)))
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
return inputs_per_image
@overload
def run_test(
hf_runner: Type[HfRunner],
@@ -103,39 +134,17 @@ def run_test(
tensor_parallel_size: int,
distributed_executor_backend: Optional[str] = None,
):
images = [asset.pil_image for asset in image_assets]
if size_factors is not None:
inputs_per_image = [(
[prompt for _ in size_factors],
[rescale_image_size(image, factor) for factor in size_factors],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
elif sizes is not None:
inputs_per_image = [(
[
prompt if size is not None else text_only_prompts[0]
for size in sizes
],
[
image.resize(size) if size is not None else None
for size in sizes
],
) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
if len(sizes) == 0:
inputs_per_image.append(
(text_only_prompts, [None] * len(text_only_prompts)))
else:
raise ValueError("You must provide either `size_factors` or `sizes`")
_run_test(hf_runner,
vllm_runner,
inputs_per_image,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend)
_run_test(
hf_runner,
vllm_runner,
_get_inputs(image_assets, size_factors=size_factors, sizes=sizes),
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
)
def _run_test(
@@ -167,8 +176,8 @@ def _run_test(
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_num_seqs=16,
max_model_len=4096,
max_num_seqs=2,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
enforce_eager=True,
@@ -185,7 +194,6 @@ def _run_test(
def process(hf_inputs: BatchEncoding):
return hf_inputs
from transformers import AutoConfig
from transformers.models.mllama import MllamaConfig as MllamaConfigHf
# use transformer's MllamaConfig for hf_runner
@@ -193,6 +201,7 @@ def _run_test(
AutoConfig.register("mllama", MllamaConfigHf, exist_ok=True)
with hf_runner(model,
dtype=dtype,
model_kwargs={"device_map": "auto"},
postprocess_inputs=process,
auto_cls=AutoModelForVision2Seq) as hf_model:
hf_outputs_per_image = [
@@ -218,26 +227,29 @@ 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.")
@pytest.mark.parametrize("model", models)
@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("sizes", SIZES)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
@@ -254,30 +266,3 @@ def test_models(hf_runner, vllm_runner, image_assets, model, sizes, dtype,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize(
"sizes",
[
[(512, 512), (1024, 512), (1536, 512), (2048, 512), (512, 1024),
(1024, 1024), (512, 1536), (512, 2028), None],
],
)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_distributed(hf_runner, vllm_runner, image_assets, model, sizes,
dtype, max_tokens, num_logprobs) -> None:
run_test(
hf_runner,
vllm_runner,
image_assets,
model,
sizes=sizes,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=2,
)