# ASUS Ascent GX10 (NVIDIA GB10) Setup Guide

**Hardware Platform:** ASUS Ascent GX10 (NVIDIA Grace Blackwell GB10)
**System Architecture：** ARM64 / NVIDIA DGX OS（Base: Ubuntu 24.04.3）

---

## 1. Environment Initialization
*This step only needs to be performed once per machine to establish a persistent development environment.*

### 1.1 Prepare Local Working Directory
To prevent source code from being lost when the container is removed, create a directory on the GX10 host machine in advance:

mkdir -p ~/my_ai_project

### 1.2 Pull the Correct Docker Image (Critical)
**Important:** You must use version 26.01 or later. Older images do not support the Blackwell architecture (`sm_121`).

docker pull nvcr.io/nvidia/pytorch:26.01-py3

### 1.3 Launch the Ultimate Container
Copy and execute the following command in full.
This command enables GPU passthrough, memory optimization, volume mounting, and network port forwarding.
* Port`8888`: Reserved for Jupyter Lab
* Port`5000`: Reserved for Flask web streaming
* `-v`: Mounts the local project directory into the container

docker run -dt \
  --name gx10_ultimate \
  --gpus all \
  --ipc=host \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  -p 8888:8888 \
  -p 5000:5000 \
  -v ~/my_ai_project:/workspace/project \
  nvcr.io/nvidia/pytorch:26.01-py3

## 2. Daily Usage Workflow
*Follow this procedure each time the system is powered on.*

### 2.1 Check Container Status

docker ps -a

### 2.2 Start and Enter the Container
If the container status is **Exited**, start it first and then enter:

docker start gx10_ultimate
docker exec -it gx10_ultimate bash

*(If the status is already **Up**, simply execute the second command.)*

## 3. Environment Configuration and Package Installation
*The following steps are executed inside the container (`root@xxxx:/workspace#`) and only need to be done once.*

### 3.1 Fix System Graphics Dependencies
(Resolves `libxcb` / OpenCV-related errors)
YOLO and OpenCV require the following low-level Linux libraries:

apt-get update && apt-get install -y \
  libgl1 \
  libglib2.0-0 \
  libsm6 \
  libxrender1 \
  libxext6 \
  libxcb1

### 3.2 Install AI-Related Python Packages
* **Ultralytics**: Includes YOLO11 / YOLOv8
* **Pycocotools**: COCO dataset utilities
* **Nvitop**: GPU performance monitoring

pip install ultralytics \
pycocotools \
nvitop \
flask \
jupyterlab

## 4. Verify GB10 Driver and Architecture Support
Ensure that PyTorch correctly detects the Blackwell architecture (`sm_120`/`sm_121`):

python -c "import torch; print(f'Arch List: {torch.cuda.get_arch_list()}')"

* **Pass:** Output contains `sm_120` or `sm_121`
* **Fail**: Output only goes up to `sm_90`
→ The Docker image is outdated; repeat Step 1.2

# 5. Performance Monitoring
It is recommended to open a second terminal window for monitoring during training.

**Option A: Official NVIDIA Tool**

watch -n 0.5 nvidia-smi

**Option B: Interactive GPU Monitor**

pip install nvitop
nvitop -m

**Option C: System Resource Monitor**

btop

# 6. Troubleshooting

| Error Message | Cause | Solution |
| ------ | ---- | --- |
| UserWarning: NVIDIA GB10... is not compatible | Outdated Docker image | Ensure you are using `26.01-py3` or newer |
| ImportError: libxcb.so.1... | Missing Linux graphics libraries | Run the `apt-get install` command in Step 3.1 |
| PIL.UnidentifiedImageError | Image download blocked (HTTP 403) | Use alternative image sources (e.g., GitHub raw links) or add a User-Agent header |