Audio-language assistant with Qwen2Audio and OpenVINO#

This Jupyter notebook can be launched after a local installation only.

Github

Qwen2-Audio is the new series of Qwen large audio-language models. Qwen2-Audio is capable of accepting various audio signal inputs and performing audio analysis or direct textual responses with regard to speech instructions. Model supports more than 8 languages and dialects, e.g., Chinese, English, Cantonese, French, Italian, Spanish, German, and Japanese and can work in two distinct audio interaction modes: * voice chat: users can freely engage in voice interactions with Qwen2-Audio without text input; * audio analysis: users could provide audio and text instructions for analysis during the interaction;

More details about model can be found in model card, blog, original repository and technical report.

In this tutorial we consider how to convert and optimize Qwen2Audio 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:

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#

%pip install -q "transformers>=4.45" "torch>=2.1" "librosa"  "gradio>=4.36" "modelscope-studio>=0.4.2" --extra-index-url https://download.pytorch.org/whl/cpu
%pip install -qU "openvino>=2024.4.0" "nncf>=2.13.0"

from pathlib import Path
import requests

if not Path("ov_qwen2_audio_helper.py").exists():
    r = requests.get(url="https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/qwen2-audio/ov_qwen2_audio_helper.py")
    open("ov_qwen2_audio_helper.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)

# Read more about telemetry collection at https://github.com/openvinotoolkit/openvino_notebooks?tab=readme-ov-file#-telemetry
from notebook_utils import collect_telemetry

collect_telemetry("qwen2-audio.ipynb")

Convert and Optimize model#

Qwen2Audio 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_audio_helper.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

Qwen2Audio 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 audio and optionally input text instruction and, that transformed to the unified embedding space using input_embedding and audio_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:

  • Audio encoder for encoding input audio into audio embedding space and Multi-modal projector for transforming audio embeddings into language model embedding space.

  • Input Embedding for conversion input text tokens into embedding space

  • Language Model for generation answer based on input embeddings provided by Audio Encoder and Input Embedding models.

Compress model weights to 4-bit#

For reducing memory consumption, weights compression optimization can be applied using NNCF.

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.

pt_model_id = "Qwen/Qwen2-Audio-7B-Instruct"

model_dir = Path(pt_model_id.split("/")[-1])

from ov_qwen2_audio_helper import convert_qwen2audio_model

# uncomment these lines to see model conversion code
# convert_qwen2audio_model??

import nncf

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

convert_qwen2audio_model(pt_model_id, model_dir, compression_configuration)

✅ Qwen/Qwen2-Audio-7B-Instruct model already converted. You can find results in Qwen2-Audio-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_audio_helper.py we defined inference class OVQwen2AudioForConditionalGeneration 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.

from ov_qwen2_audio_helper import OVQwen2AudioForConditionalGeneration

# Uncomment below lines to see the model inference class code
# OVQwen2AudioForConditionalGeneration??

from notebook_utils import device_widget

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

device

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

ov_model = OVQwen2AudioForConditionalGeneration(model_dir, device.value)

Run model inference#

from transformers import AutoProcessor, TextStreamer
import librosa
import IPython.display as ipd


processor = AutoProcessor.from_pretrained(model_dir)

audio_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/1272-128104-0000.flac"
audio_chat_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen2-Audio/audio/guess_age_gender.wav"
audio_file = Path(audio_url.split("/")[-1])
audio_chat_file = Path(audio_chat_url.split("/")[-1])

if not audio_file.exists():
    r = requests.get(audio_url)
    with audio_file.open("wb") as f:
        f.write(r.content)

if not audio_chat_file.exists():
    r = requests.get(audio_chat_url)
    with audio_chat_file.open("wb") as f:
        f.write(r.content)

Voice chat#

conversation = [
    {"role": "system", "content": "You are a helpful assistant."},
    {
        "role": "user",
        "content": [
            {"type": "audio", "audio_url": audio_chat_url},
        ],
    },
]

text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
audios = [librosa.load(audio_chat_file, sr=processor.feature_extractor.sampling_rate)[0]]

inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
display(ipd.Audio(audio_chat_file))
print("Answer:")

generate_ids = ov_model.generate(**inputs, max_new_tokens=50, streamer=TextStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True))

It is strongly recommended to pass the sampling_rate argument to this function. Failing to do so can result in silent errors that might be hard to debug.
Setting pad_token_id to eos_token_id:151645 for open-end generation.
Answer:
Yes, I can guess that you are a female in your twenties.

Audio analysis#

question = "What does the person say?"

conversation = [
    {"role": "system", "content": "You are a helpful assistant."},
    {
        "role": "user",
        "content": [
            {"type": "audio", "audio_url": audio_url},
            {"type": "text", "text": question},
        ],
    },
]

text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
audios = [librosa.load(audio_file, sr=processor.feature_extractor.sampling_rate)[0]]

inputs = processor(text=text, audios=audios, return_tensors="pt", padding=True)
print("Question:")
print(question)
display(ipd.Audio(audio_file))
print("Answer:")

generate_ids = ov_model.generate(**inputs, max_new_tokens=50, streamer=TextStreamer(processor.tokenizer, skip_prompt=True, skip_special_tokens=True))

It is strongly recommended to pass the sampling_rate argument to this function. Failing to do so can result in silent errors that might be hard to debug.
Question:
What does the person say?

Setting pad_token_id to eos_token_id:151645 for open-end generation.
Answer:
The person says: 'Mister Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'

Interactive Demo#

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)

from gradio_helper import make_demo


demo = make_demo(ov_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/