Notebook

Visual-language assistant with Qwen2VL and OpenVINO¶

Qwen2VL is the latest addition to the QwenVL series of multimodal large language models.

Key Enhancements of Qwen2VL:

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 Details:

Naive Dynamic Resolution: 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.

More details about model can be found in model card, blog and original repo.

In this tutorial we consider how to convert and optimize Qwen2VL model for creating multimodal chatbot. Additionally, we demonstrate how to apply stateful transformation on LLM part and model optimization techniques like weights compression using NNCF

Table of contents:¶

Prerequisites
Select model
Convert and Optimize model
- Compress model weights to 4-bit
Prepare model inference pipeline
- Select inference device
Run model inference
Interactive Demo

Installation Instructions¶

This is a self-contained example that relies solely on its own code.

We recommend running the notebook in a virtual environment. You only need a Jupyter server to start. For details, please refer to Installation Guide.

Prerequisites¶

In [1]:

%pip install -q "transformers>=4.45" "torch>=2.1" "torchvision" "qwen-vl-utils" "Pillow" "gradio>=4.36" --extra-index-url https://download.pytorch.org/whl/cpu
%pip install -qU "openvino>=2024.4.0" "nncf>=2.13.0"

In [2]:

from pathlib import Path
import requests

if not Path("ov_qwen2_vl.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/qwen2-vl/ov_qwen2_vl.py")
    open("ov_qwen2_vl.py", "w").write(r.text)

if not Path("notebook_utils.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/notebook_utils.py")
    open("notebook_utils.py", "w").write(r.text)

Select model¶

There are multiple Qwen2VL models available in models collection. You can select one of them for conversion and optimization in notebook using widget bellow:

In [3]:

from ov_qwen2_vl import model_selector

model_id = model_selector()

model_id

INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, tensorflow, onnx, openvino

2024-10-30 21:02:36.765098: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2024-10-30 21:02:36.777073: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:477] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
E0000 00:00:1730307756.791531  559916 cuda_dnn.cc:8310] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
E0000 00:00:1730307756.795971  559916 cuda_blas.cc:1418] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2024-10-30 21:02:36.810854: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.

Out[3]:

Dropdown(description='Model:', options=('Qwen/Qwen2-VL-2B-Instruct', 'Qwen/Qwen2-VL-7B-Instruct'), value='Qwen…

In [4]:

print(f"Selected {model_id.value}")
pt_model_id = model_id.value
model_dir = Path(pt_model_id.split("/")[-1])

Selected Qwen/Qwen2-VL-7B-Instruct

Convert and Optimize model¶

Qwen2VL is PyTorch model. OpenVINO supports PyTorch models via conversion to OpenVINO Intermediate Representation (IR). OpenVINO model conversion API should be used for these purposes. ov.convert_model function accepts original PyTorch model instance and example input for tracing and returns ov.Model representing this model in OpenVINO framework. Converted model can be used for saving on disk using ov.save_model function or directly loading on device using core.compile_model. ov_qwen2_vl.py script contains helper function for model conversion, please check its content if you interested in conversion details.

Click here for more detailed explanation of conversion steps

Qwen2VL is autoregressive transformer generative model, it means that each next model step depends from model output from previous step. The generation approach is based on the assumption that the probability distribution of a word sequence can be decomposed into the product of conditional next word distributions. In other words, model predicts the next token in the loop guided by previously generated tokens until the stop-condition will be not reached (generated sequence of maximum length or end of string token obtained). The way the next token will be selected over predicted probabilities is driven by the selected decoding methodology. You can find more information about the most popular decoding methods in this blog. The entry point for the generation process for models from the Hugging Face Transformers library is the `generate` method. You can find more information about its parameters and configuration in the documentation. To preserve flexibility in the selection decoding methodology, we will convert only model inference for one step.

The inference flow has difference on first step and for the next. On the first step, model accept preprocessed input instruction and image, that transformed to the unified embedding space using input_embedding and image_encoder models, after that language model, LLM-based part of model, runs on input embeddings to predict probability of next generated tokens. On the next step, language_model accepts only next token id selected based on sampling strategy and processed by input_embedding model and cached attention key and values. Since the output side is auto-regressive, an output token hidden state remains the same once computed for every further generation step. Therefore, recomputing it every time you want to generate a new token seems wasteful. With the cache, the model saves the hidden state once it has been computed. The model only computes the one for the most recently generated output token at each time step, re-using the saved ones for hidden tokens. This reduces the generation complexity from $O(n^3)$ to $O(n^2)$ for a transformer model. More details about how it works can be found in this article. To sum up above, model consists of 4 parts:

Image encoder for encoding input images into embedding space.
Input Embedding for conversion input text tokens into embedding space
Language Model for generation answer based on input embeddings provided by Image Encoder and Input Embedding models.

Compress model weights to 4-bit¶

Click here for more details about weight compression

Weight compression aims to reduce the memory footprint of a model. It can also lead to significant performance improvement for large memory-bound models, such as Large Language Models (LLMs). LLMs and other models, which require extensive memory to store the weights during inference, can benefit from weight compression in the following ways:

enabling the inference of exceptionally large models that cannot be accommodated in the memory of the device;
improving the inference performance of the models by reducing the latency of the memory access when computing the operations with weights, for example, Linear layers.

Neural Network Compression Framework (NNCF) provides 4-bit / 8-bit mixed weight quantization as a compression method primarily designed to optimize LLMs. The main difference between weights compression and full model quantization (post-training quantization) is that activations remain floating-point in the case of weights compression which leads to a better accuracy. Weight compression for LLMs provides a solid inference performance improvement which is on par with the performance of the full model quantization. In addition, weight compression is data-free and does not require a calibration dataset, making it easy to use.

nncf.compress_weights function can be used for performing weights compression. The function accepts an OpenVINO model and other compression parameters. Compared to INT8 compression, INT4 compression improves performance even more, but introduces a minor drop in prediction quality.

More details about weights compression, can be found in OpenVINO documentation.

In [5]:

from ov_qwen2_vl import convert_qwen2vl_model

# uncomment these lines to see model conversion code
# convert_qwen2vl_model??

In [6]:

import nncf

compression_configuration = {
    "mode": nncf.CompressWeightsMode.INT4_ASYM,
    "group_size": 128,
    "ratio": 1.0,
}

convert_qwen2vl_model(pt_model_id, model_dir, compression_configuration)

⌛ Qwen/Qwen2-VL-7B-Instruct conversion started. Be patient, it may takes some time.
⌛ Load Original model

`Qwen2VLRotaryEmbedding` can now be fully parameterized by passing the model config through the `config` argument. All other arguments will be removed in v4.46

Loading checkpoint shards:   0%|          | 0/5 [00:00<?, ?it/s]

✅ Original model successfully loaded
⌛ Convert Input embedding model
WARNING:nncf:NNCF provides best results with torch==2.4.*, while current torch version is 2.5.1+cu124. If you encounter issues, consider switching to torch==2.4.*
✅ Input embedding model successfully converted
⌛ Convert Language model

/home/ea/work/py311/lib/python3.11/site-packages/transformers/modeling_utils.py:4779: FutureWarning: `_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead
  warnings.warn(
/home/ea/work/py311/lib/python3.11/site-packages/transformers/cache_utils.py:447: TracerWarning: Using len to get tensor shape might cause the trace to be incorrect. Recommended usage would be tensor.shape[0]. Passing a tensor of different shape might lead to errors or silently give incorrect results.
  or len(self.key_cache[layer_idx]) == 0  # the layer has no cache
/home/ea/work/py311/lib/python3.11/site-packages/transformers/models/qwen2_vl/modeling_qwen2_vl.py:477: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
  if sequence_length != 1:
/home/ea/work/py311/lib/python3.11/site-packages/transformers/cache_utils.py:432: TracerWarning: Using len to get tensor shape might cause the trace to be incorrect. Recommended usage would be tensor.shape[0]. Passing a tensor of different shape might lead to errors or silently give incorrect results.
  elif len(self.key_cache[layer_idx]) == 0:  # fills previously skipped layers; checking for tensor causes errors

✅ Language model successfully converted
⌛ Weights compression with int4_asym mode started
INFO:nncf:Statistics of the bitwidth distribution:
┍━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┑
│   Num bits (N) │ % all parameters (layers)   │ % ratio-defining parameters (layers)   │
┝━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┥
│              8 │ 8% (1 / 197)                │ 0% (0 / 196)                           │
├────────────────┼─────────────────────────────┼────────────────────────────────────────┤
│              4 │ 92% (196 / 197)             │ 100% (196 / 196)                       │
┕━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┙

Output()

✅ Weights compression finished
⌛ Convert Image embedding model
⌛ Weights compression with int4_asym mode started
INFO:nncf:Statistics of the bitwidth distribution:
┍━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┑
│   Num bits (N) │ % all parameters (layers)   │ % ratio-defining parameters (layers)   │
┝━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┥
│              8 │ 3% (1 / 130)                │ 0% (0 / 129)                           │
├────────────────┼─────────────────────────────┼────────────────────────────────────────┤
│              4 │ 97% (129 / 130)             │ 100% (129 / 129)                       │
┕━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┙

Output()

✅ Weights compression finished
✅ Image embedding model successfully converted
✅ Qwen/Qwen2-VL-7B-Instruct model conversion finished. You can find results in Qwen2-VL-7B-Instruct

Prepare model inference pipeline¶

As discussed, the model comprises Image Encoder and LLM (with separated text embedding part) that generates answer. In ov_qwen2_vl.py we defined inference class OVQwen2VLModel that will represent generation cycle, It is based on HuggingFace Transformers GenerationMixin and looks similar to Optimum Intel OVModelForCausalLM that is used for LLM inference.

In [7]:

from ov_qwen2_vl import OVQwen2VLModel

# Uncomment below lines to see the model inference class code
# OVQwen2VLModel??

Select inference device¶

In [8]:

from notebook_utils import device_widget

device = device_widget(default="AUTO", exclude=["NPU"])

device

Out[8]:

Dropdown(description='Device:', index=1, options=('CPU', 'AUTO'), value='AUTO')

In [9]:

model = OVQwen2VLModel(model_dir, device.value)

Run model inference¶

In [10]:

from PIL import Image
from transformers import AutoProcessor, AutoTokenizer
from qwen_vl_utils import process_vision_info
from transformers import TextStreamer


min_pixels = 256 * 28 * 28
max_pixels = 1280 * 28 * 28
processor = AutoProcessor.from_pretrained(model_dir, min_pixels=min_pixels, max_pixels=max_pixels)

if processor.chat_template is None:
    tok = AutoTokenizer.from_pretrained(model_dir)
    processor.chat_template = tok.chat_template

example_image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
example_image_path = Path("demo.jpeg")

if not example_image_path.exists():
    Image.open(requests.get(example_image_url, stream=True).raw).save(example_image_path)

image = Image.open(example_image_path)
question = "Describe this image."

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": f"file://{example_image_path}",
            },
            {"type": "text", "text": question},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)

display(image)
print("Question:")
print(question)
print("Answer:")

generated_ids = model.generate(**inputs, max_new_tokens=100, streamer=TextStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True))

Setting `pad_token_id` to `eos_token_id`:None for open-end generation.

Question:
Describe this image.
Answer:
The image depicts a serene beach scene at sunset. A woman and her dog are sitting on the sandy shore, enjoying each other's company. The woman is wearing a plaid shirt and has long hair. She is holding the dog's paw, and the dog is wearing a colorful harness. The dog appears to be a large breed, possibly a Labrador Retriever. The ocean is visible in the background, with gentle waves and a clear sky. The sun is setting, casting a warm glow over

In [11]:

if not Path("gradio_helper.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/qwen2-vl/gradio_helper.py")
    open("gradio_helper.py", "w").write(r.text)

Interactive Demo¶

Now, you can try to chat with model. Upload image or video using Upload button, provide your text message into Input field and click Submit to start communication.

In [ ]:

from gradio_helper import make_demo


demo = make_demo(model, processor)

try:
    demo.launch(debug=True)
except Exception:
    demo.launch(debug=True, share=True)
# if you are launching remotely, specify server_name and server_port
# demo.launch(server_name='your server name', server_port='server port in int')
# Read more in the docs: https://gradio.app/docs/