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

[Model] Make llama3.2 support multiple and interleaved images (#9095)

Browse Source
This commit is contained in:
Xiang Xu
2024-10-14 15:24:26 -07:00
committed by GitHub
parent 4d31cd424b
commit f0fe4fe86d
3 changed files with 384 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
tests/models/encoder_decoder/vision_language/test_mllama.py
View File

@@ -12,7 +12,7 @@ from ....conftest import (IMAGE_ASSETS, HfRunner, PromptImageInput, VllmRunner,
from ....utils import large_gpu_test
from ...utils import check_logprobs_close
_LIMIT_IMAGE_PER_PROMPT = 1
_LIMIT_IMAGE_PER_PROMPT = 3
HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
"stop_sign":
@@ -244,8 +244,9 @@ def _run_test(
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(hf_runner, vllm_runner, image_assets, model, sizes, dtype,
max_tokens, num_logprobs) -> None:
def test_models_single_leading_image(hf_runner, vllm_runner, image_assets,
model, sizes, dtype, max_tokens,
num_logprobs) -> None:
run_test(
hf_runner,
vllm_runner,
@@ -257,3 +258,81 @@ def test_models(hf_runner, vllm_runner, image_assets, model, sizes, dtype,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_multi_leading_images(hf_runner, vllm_runner, image_assets,
model, dtype, max_tokens,
num_logprobs) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"<|image|><|image|><|begin_of_text|>Describe 2 images.", # noqa: E501
"<|image|><|image|><|begin_of_text|>Describe 2 images.", # noqa: E501
"<|image|><|image|><|image|><|begin_of_text|>Describe 3 images.", # noqa: E501
],
[
[stop_sign, cherry_blossom],
# Images with different sizes.
[
stop_sign.resize((512, 512)),
stop_sign,
],
[
stop_sign,
stop_sign.resize((512, 1536)),
cherry_blossom.resize((512, 1024)),
],
])]
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)
@large_gpu_test(min_gb=48)
@pytest.mark.parametrize("model", models)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [128])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models_interleaved_images(hf_runner, vllm_runner, image_assets, model,
dtype, max_tokens, num_logprobs) -> None:
stop_sign = image_assets[0].pil_image
cherry_blossom = image_assets[1].pil_image
inputs = [(
[
"<|begin_of_text|>The content of the image <|image|> is", # noqa: E501
"<|begin_of_text|>Between the first image <|image|> and the second image<|image|>, " # noqa: E501
"which is a stop sign and which is a cherry blossom?", # noqa: E501
],
[
[stop_sign],
[stop_sign, cherry_blossom],
])]
_run_test(
hf_runner,
vllm_runner,
inputs,
model,
dtype=dtype,
max_tokens=max_tokens,
num_logprobs=num_logprobs,
tensor_parallel_size=1,
)