Qwen2-VL

Introduction

We're excited to unveil Qwen2-VL, the latest iteration of our Qwen-VL model, representing nearly a year of innovation.

What’s New in Qwen2-VL?

Key Enhancements:

• SoTA understanding of images of various resolution & ratio: Qwen2-VL achieves state-of-the-art performance on visual understanding benchmarks, including MathVista, DocVQA, RealWorldQA, MTVQA, etc.

• Understanding videos of 20min+: Qwen2-VL can understand videos over 20 minutes for high-quality video-based question answering, dialog, content creation, etc.

• Agent that can operate your mobiles, robots, etc.: with the abilities of complex reasoning and decision making, Qwen2-VL can be integrated with devices like mobile phones, robots, etc., for automatic operation based on visual environment and text instructions.

• Multilingual Support: to serve global users, besides English and Chinese, Qwen2-VL now supports the understanding of texts in different languages inside images, including most European languages, Japanese, Korean, Arabic, Vietnamese, etc.

Model Architecture Updates:

• Naive Dynamic Resolution: Unlike before, Qwen2-VL can handle arbitrary image resolutions, mapping them into a dynamic number of visual tokens, offering a more human-like visual processing experience.

• Multimodal Rotary Position Embedding (M-ROPE): Decomposes positional embedding into parts to capture 1D textual, 2D visual, and 3D video positional information, enhancing its multimodal processing capabilities.

We have three models with 2, 7 and 72 billion parameters. This repo contains the instruction-tuned 7B Qwen2-VL model. For more information, visit our Blog and GitHub.

Benchmark

Performance of Quantized Models

This section reports the generation performance of quantized models (including GPTQ and AWQ) of the Qwen2-VL series. Specifically, we report:

• MMMU_VAL (Accuracy)
• DocVQA_VAL (Accuracy)
• MMBench_DEV_EN (Accuracy)
• MathVista_MINI (Accuracy)

We use VLMEvalkit to evaluate all models.

Speed Benchmark

This section reports the speed performance of bf16 models, quantized models (including GPTQ-Int4, GPTQ-Int8 and AWQ) of the Qwen2-VL series. Specifically, we report the inference speed (tokens/s) as well as memory footprint (GB) under the conditions of different context lengths.

The environment of the evaluation with huggingface transformers is:

• NVIDIA A100 80GB
• CUDA 11.8
• Pytorch 2.2.1+cu118
• Flash Attention 2.6.1
• Transformers 4.38.2
• AutoGPTQ 0.6.0+cu118
• AutoAWQ 0.2.5+cu118 (autoawq_kernels 0.0.6+cu118)

Note:

• We use the batch size of 1 and the least number of GPUs as possible for the evalution.

• We test the speed and memory of generating 2048 tokens with the input lengths of 1, 6144, 14336, 30720, 63488, and 129024 tokens (>32k is only avaliable for Qwen2-72B-Instuct and Qwen2-7B-Instuct).

• 7B (transformers)

Requirements

The code of Qwen2-VL has been in the latest Hugging face transformers and we advise you to build from source with command pip install git+https://github.com/huggingface/transformers, or you might encounter the following error:

KeyError: 'qwen2_vl'

Quickstart

We offer a toolkit to help you handle various types of visual input more conveniently. This includes base64, URLs, and interleaved images and videos. You can install it using the following command:

pip install qwen-vl-utils

Here we show a code snippet to show you how to use the chat model with transformers and qwen_vl_utils:

from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

# default: Load the model on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-7B-Instruct-AWQ", torch_dtype="auto", device_map="auto"
)

# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen2VLForConditionalGeneration.from_pretrained(
#     "Qwen/Qwen2-VL-7B-Instruct-AWQ",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# default processer
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-