Qwen2.5-VL-3B-Instruct

This version of Qwen2.5-VL-3B-Instruct has been converted to run on the Axera NPU using w8a16 quantization.

This model has been optimized with the following LoRA:

Compatible with Pulsar2 version: 3.4

Convert tools links:

For those who are interested in model conversion, you can try to export axmodel through the original repo : https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct

Pulsar2 Link, How to Convert LLM from Huggingface to axmodel

AXera NPU HOST LLM Runtime

Support Platform

• AX650
  • AX650N DEMO Board
  • M4N-Dock(爱芯派Pro)
  • M.2 Accelerator card

Image Process

Chips	input size	image num	image encoder	ttft(320 tokens)	w8a16	DDR	Flash
AX650	448*448	1	780 ms	2857 ms	6.2 tokens/sec	4.3 GiB	4.6 GiB

Video Process

Chips	input size	image num	image encoder	ttft(512 tokens)	w8a16	DDR	Flash
AX650	308*308	8	1400 ms	5400 ms	6.1 tokens/sec	4.4 GiB	4.7 GiB

The DDR capacity refers to the CMM memory that needs to be consumed. Ensure that the CMM memory allocation on the development board is greater than this value.

How to use

Download all files from this repository to the device

If you using AX650 Board

root@ax650:/mnt/qtang/llm-test/qwen2.5-vl-3b# tree -L 2
.
├── image
│   └── ssd_car.jpg
├── main
├── python
│   ├── cv_resize.py
│   ├── infer_image.py
│   ├── infer_text.py
│   ├── infer_video.py
│   ├── preprocess.py
│   └── utils.py
├── qwen2_5-vl-3b-image-ax650
│   ├── Qwen2.5-VL-3B-Instruct_vision_nchw448.axmodel
│   ├── model.embed_tokens.weight.bfloat16.bin
│   ├── qwen2_5_vl_p320_l0_together.axmodel
......
│   ├── qwen2_5_vl_p320_l9_together.axmodel
│   └── qwen2_5_vl_post.axmodel
├── qwen2_5-vl-3b-video-ax650
│   ├── Qwen2.5-VL-3B-Instruct_vision_nhwc.axmodel
│   ├── model.embed_tokens.weight.bfloat16.bin
│   ├── qwen2_5_vl_p512_l0_together.axmodel
......
│   ├── qwen2_5_vl_p512_l9_together.axmodel
│   └── qwen2_5_vl_post.axmodel
├── qwen2_5-vl-tokenizer
│   ├── chat_template.json
│   ├── config.json
│   ├── generation_config.json
│   ├── merges.txt
│   ├── model.safetensors.index.json
│   ├── preprocessor_config.json
│   ├── tokenizer.json
│   ├── tokenizer_config.json
│   └── vocab.json
├── qwen2_tokenizer_image_448.py
├── qwen2_tokenizer_video_308.py
├── run_qwen2_5_vl_image.sh
├── run_qwen2_5_vl_video.sh
└── video
    ├── frame_0075.jpg
......
    └── frame_0089.jpg

Prepare tokenizer server

Install transformer

pip install transformers==4.41.1

Demo Run

Image understand demo

start tokenizer server for image understand demo

python3 qwen2_tokenizer_image_448.py --port 12345

run image understand demo

• input text

描述下图片

• input image

root@ax650:/mnt/qtang/llm-test/qwen2.5-vl-3b# ./run_qwen2_5_vl_image.sh
[I][                            Init][ 129]: LLM init start
bos_id: -1, eos_id: 151645
  2% | █                                 |   1 /  40 [0.01s<0.24s, 166.67 count/s] tokenizer init ok
[I][                            Init][  26]: LLaMaEmbedSelector use mmap
100% | ████████████████████████████████ |  40 /  40 [38.23s<38.23s, 1.05 count/s] init vpm axmodel ok,remain_cmm(7600 MB)
[I][                            Init][ 277]: max_token_len : 1023
[I][                            Init][ 282]: kv_cache_size : 256, kv_cache_num: 1023
[I][                            Init][ 290]: prefill_token_num : 320
[I][                            Init][ 292]: vpm_height : 1024,vpm_width : 392
[I][                            Init][ 301]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running

prompt >> who are you?
image >>
[I][                             Run][ 638]: ttft: 2854.47 ms
I am a large language model created by Alibaba Cloud. I am called Qwen.

[N][                             Run][ 779]: hit eos,avg 6.05 token/s

prompt >> 描述下图片
image >> image/ssd_car.jpg
[I][                          Encode][ 416]: image encode time : 795.614014 ms, size : 524288
[I][                             Run][ 638]: ttft: 2856.88 ms
这张图片展示了一条繁忙的城市街道。前景中，一名女子站在人行道上，她穿着黑色外套，面带微笑。她旁边是一辆红色的双层巴士，巴士上有一个广告，
上面写着“THINGS GET MORE EXITING WHEN YOU SAY ‘YES’”。巴士的车牌号是“L15”。巴士旁边停着一辆黑色的小型货车。背景中可以看到一些商店和行人，
街道两旁的建筑物是现代的玻璃幕墙建筑。整体氛围显得繁忙而充满活力。

[N][                             Run][ 779]: hit eos,avg 5.96 token/s

Video understand demo

Please pre-process the image of the video file into a 308x308 size picture

start tokenizer server for image understand demo

python qwen2_tokenizer_video_308.py --port 12345

run image understand demo

root@ax650:/mnt/qtang/llm-test/qwen2.5-vl-3b# ./run_qwen2_5_vl_video.sh
[I][                            Init][ 129]: LLM init start
bos_id: -1, eos_id: 151645
  2% | █                                 |   1 /  40 [0.00s<0.12s, 333.33 count/s] tokenizer init ok
[I][                            Init][  26]: LLaMaEmbedSelector use mmap
100% | ████████████████████████████████ |  40 /  40 [40.05s<40.05s, 1.00 count/s] init vpm axmodel ok,remain_cmm(7680 MB)
[I][                            Init][ 277]: max_token_len : 1023
[I][                            Init][ 282]: kv_cache_size : 256, kv_cache_num: 1023
[I][                            Init][ 290]: prefill_token_num : 512
[I][                            Init][ 292]: vpm_height : 484,vpm_width : 392
[I][                            Init][ 301]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running

prompt >> 描述下视频
image >> video
video/frame_0000.jpg
video/frame_0008.jpg
video/frame_0016.jpg
video/frame_0024.jpg
video/frame_0032.jpg
video/frame_0040.jpg
video/frame_0048.jpg
video/frame_0056.jpg
[I][                          Encode][ 416]: image encode time : 1487.557007 ms, size : 991232
[I][                             Run][ 638]: ttft: 5488.29 ms
视频展示了两只松鼠在户外的场景。背景是模糊的山脉和蓝天，前景中有松鼠在互动。松鼠的毛色主要是棕色和白色，它们的爪子是橙色的。松鼠似乎在互相玩耍或争抢，它们的爪子和嘴巴都伸向对方。整个场景显得非常自然和生动。

Inference with M.2 Accelerator card

What is M.2 Accelerator card?, Show this DEMO based on Raspberry PI 5.

TODO