Vision Language Model

Introduction

This section describes how to experience on-device Vision Language Model (VLM) on the RDK platform. Thanks to the excellent results of InternVL and SmolVLM, we have achieved quantization and deployment on the RDK platform. This demo leverages the powerful KV Cache management in llama.cpp, combined with the computational advantages of the RDK platform's BPU module, to enable local VLM model deployment.

Code repository: (https://github.com/D-Robotics/hobot_llamacpp.git)

Supported Platforms

Platform	OS / Method	Demo Functionality
RDK X5, RDK X5 Module	Ubuntu 22.04 (Humble)	On-device Vision Language Model
RDK S100, RDK S100P	Ubuntu 22.04 (Humble)	On-device Vision Language Model

Supported Models

Model Name	Para	Platform	Image Encoder	Language Encoder and Decoder
InternVL2_5	1B	X5	vit_model_int16_v2.bin	Qwen2.5-0.5B-Instruct-Q4_0.gguf
InternVL2_5	1B	S100	vit_model_int16.hbm	Qwen2.5-0.5B-Instruct-Q4_0.gguf
InternVL3	1B	X5	vit_model_int16_VL3_1B_Instruct_X5.bin	qwen2_5_q8_0_InternVL3_1B_Instruct.gguf
InternVL3	1B	S100	vit_model_int16_VL3_1B_Instruct.hbm	qwen2_5_q8_0_InternVL3_1B_Instruct.gguf
InternVL3	2B	X5	vit_model_int16_VL3_2B_Instruct.bin	qwen2_5_1.5b_q8_0_InternVL3_2B_Instruct.gguf
InternVL3	2B	S100	vit_model_int16_VL3_2B_Instruct.hbm	qwen2_5_1.5b_q8_0_InternVL3_2B_Instruct.gguf
SmolVLM2	256M	X5	SigLip_int16_SmolVLM2_256M_Instruct_MLP_C1_UP_X5.bin	SmolVLM2-256M-Video-Instruct-Q8_0.gguf
SmolVLM2	256M	S100	SigLip_int16_SmolVLM2_256M_Instruct_S100.hbm	SmolVLM2-256M-Video-Instruct-Q8_0.gguf
SmolVLM2	500M	X5	SigLip_int16_SmolVLM2_500M_Instruct_MLP_C1_UP_X5.bin	SmolVLM2-500M-Video-Instruct-Q8_0.gguf
SmolVLM2	500M	S100	SigLip_int16_SmolVLM2_500M_Instruct_S100.hbm	SmolVLM2-500M-Video-Instruct-Q8_0.gguf

Preparation

RDK Platform

1. RDK should be flashed with the Ubuntu 22.04 system image.
2. TogetheROS.Bot must be successfully installed on the RDK.
3. Install package

sudo apt update
sudo apt install tros-humble-hobot-llamacpp

4. System Setup

Use command srpi-config for setting ION memory larger than 1.6GB, then restart and setup the maximum CPU frequency.

• For RDK X5:

  1. srpi-configUsage Performance Options

  2. CPU Frequency: CPU Frequency Setting

• For RDK S100:

  1. srpi-config Usage Performance Options

  2. CPU Frequency: CPU Frequency Setting

Usage

RDK Platform

Use the srpi-config command to set the ION memory size to 1.6GB. For details, refer to the Performance Options section in the RDK User Manual.

After rebooting, set the CPU maximum frequency to 1.5GHz and the scheduling mode to performance with the following commands:

sudo bash -c 'echo 1 > /sys/devices/system/cpu/cpufreq/boost'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu0/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu1/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu2/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu3/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu4/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu5/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu6/cpufreq/scaling_governor'
sudo bash -c 'echo performance >/sys/devices/system/cpu/cpu7/cpufreq/scaling_governor'

Currently, two demo modes are provided: direct terminal input (image and text), or subscribing to image and text messages and publishing the results as text.

InternVL

Before running the program, download the model files to the working directory with the following commands:

# Download model files
wget https://huggingface.co/D-Robotics/InternVL2_5-1B-GGUF-BPU/blob/main/Qwen2.5-0.5B-Instruct-Q4_0.gguf
wget https://huggingface.co/D-Robotics/InternVL2_5-1B-GGUF-BPU/blob/main/rdkx5/vit_model_int16_v2.bin

source /opt/tros/humble/setup.bash
cp -r /opt/tros/${TROS_DISTRO}/lib/hobot_llamacpp/config/ .

ros2 run hobot_llamacpp hobot_llamacpp --ros-args -p feed_type:=0 -p image:=config/image2.jpg -p image_type:=0 -p user_prompt:="Describe the image." -p model_file_name:=vit_model_int16_v2.bin -p llm_model_name:=Qwen2.5-0.5B-Instruct-Q4_0.gguf

After starting the program, you can use a local image and custom prompt for output.

SmolVLM

Before running the program, download the model files to the working directory with the following commands:

# Download model files
wget https://huggingface.co/D-Robotics/SmolVLM2-256M-Video-Instruct-GGUF-BPU/resolve/main/rdkx5/SigLip_int16_SmolVLM2_256M_Instruct_MLP_C1_UP_X5.bin
wget https://huggingface.co/D-Robotics/SmolVLM2-256M-Video-Instruct-GGUF-BPU/resolve/main/SmolVLM2-256M-Video-Instruct-Q8_0.gguf

source /opt/tros/humble/setup.bash
cp -r /opt/tros/${TROS_DISTRO}/lib/hobot_llamacpp/config/ .
ros2 run hobot_llamacpp hobot_llamacpp --ros-args -p feed_type:=0 -p model_type:=1 -p image:=config/image2.jpg -p image_type:=0 -p user_prompt:="Describe the image." -p model_file_name:=SigLip_int16_SmolVLM2_256M_Instruct_MLP_C1_UP_X5.bin -p llm_model_name:=SmolVLM2-256M-Video-Instruct-Q8_0.gguf

After starting the program, you can use a local image and custom prompt for output.

Notes

Ensure the development board more than 1.6GB size of ION memory, otherwise the model may fail to load.