A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 18:15 UTC
[Bugfix] Fix Phi-3v crash when input images are of certain sizes (#7840)
zifeitong committed
Aug 24, 2024 at 18:16 UTC
80162c44b1d1e59a2c10f65b6adb9b0407439b1f
2 files changed
+22
-7
tests/models/test_phi3v.py
+22
-5
@@ -3,13 +3,14 @@ import re
3
from typing import List, Optional, Tuple, Type
4
5
import pytest
6
+from PIL import Image
7
from transformers import AutoTokenizer
8
9
from vllm.multimodal.utils import rescale_image_size
10
from vllm.sequence import SampleLogprobs
11
from vllm.utils import is_cpu, is_hip
12
12
-from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
13
+from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner
14
from .utils import check_logprobs_close
15
16
pytestmark = pytest.mark.vlm
@@ -58,7 +59,7 @@ if is_hip():
59
def run_test(
60
hf_runner: Type[HfRunner],
61
vllm_runner: Type[VllmRunner],
61
- image_assets: _ImageAssets,
62
+ images: List[Image.Image],
63
model: str,
64
*,
65
size_factors: List[float],
@@ -77,8 +78,6 @@ def run_test(
78
Note, the text input is also adjusted to abide by vllm contract.
79
The text output is sanitized to be able to compare with hf.
80
"""
80
- images = [asset.pil_image for asset in image_assets]
81
-
81
inputs_per_image = [(
82
[prompt for _ in size_factors],
83
[
@@ -159,7 +158,7 @@ def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
158
run_test(
159
hf_runner,
160
vllm_runner,
162
- image_assets,
161
+ [asset.pil_image for asset in image_assets],
162
model,
163
size_factors=size_factors,
164
dtype=dtype,
@@ -167,3 +166,21 @@ def test_models(hf_runner, vllm_runner, image_assets, model, size_factors,
166
num_logprobs=num_logprobs,
167
tensor_parallel_size=1,
168
)
169
+
170
+
171
+@pytest.mark.parametrize("model", models)
172
+@pytest.mark.parametrize("dtype", [target_dtype])
173
+def test_regression_7840(hf_runner, vllm_runner, image_assets, model,
174
+ dtype) -> None:
175
+ # Regression test for #7840.
176
+ run_test(
177
+ hf_runner,
178
+ vllm_runner,
179
+ [image_assets[0].pil_image.resize((465, 226))],
180
+ model,
181
+ size_factors=[1.0],
182
+ dtype=dtype,
183
+ max_tokens=128,
184
+ num_logprobs=10,
185
+ tensor_parallel_size=1,
186
+ )
vllm/model_executor/models/phi3v.py
-2
@@ -400,8 +400,6 @@ def input_processor_for_phi3v(ctx: InputContext, llm_inputs: LLMInputs):
400
image_data = multi_modal_data["image"]
401
if isinstance(image_data, Image.Image):
402
w, h = image_data.size
403
- w, h = _calc_hd_transform_size(width=w, height=h)
404
-
403
image_feature_size = get_phi3v_image_feature_size(hf_config,
404
input_width=w,
405
input_height=h)