Workload Registration Example

This document provides examples of workload registration for deep learning development
For detailed instructions on the actual registration process, please refer to the Register New Workload documentation.

Overview

The koojy717/lora-tuning-blackwell:1.01 image contains essential frameworks and tools required for deep learning development.

Key Components

• PyTorch 2.7.0 (Supports CUDA 12.8)
• TensorFlow 2.12.0
• Transformers 4.28.0
• PEFT 0.3.0
• Jupyter Notebook/Lab environment
• System Management Tools (net-tools, ping, traceroute)

※ Note: PyTorch version 2.7.0 is a nightly build rapidly developed for NVIDIA’s Blackwell architecture; as such, it may not exhibit consistent performance.

If you prefer a more stable, previous version of PyTorch, please use thekoojy717/lora-tuning:1.02 image, which includes Pytorch 2.01(CUDA 11.8 지원).

Workload Configuration Guide

Container Image

=== "GitHub"
    ```
    data-alliance/lora-tuning-blackwell:1.01
    ```

=== "Docker Hub"
    ```
    koojy717/lora-tuning-blackwell:1.01
    ```

=== "Lower Pytorch Version Docker Hub"
        ```
        koojy717/lora-tuning:1.02
        ```

---

Target Specifications

Recommended settings based on workload scale:

Target	Tier	GPU Memory	Usage
Large-scale Training	Tier 1 (A100, H100)	40GB+	Large-scale model training, requiring high performance.
Mid-scale Experiment	Tier 2 (전용 서버)	20GB	General model development and mid-scale experiments.
Dev / Testing	Tier 3 (PC방, 개인)	10GB	Code development and small-scale experiments.

⚡ For detailed instructions on how to register a workload, please refer to the Register New Workload page.

---

Getting Started with Jupyter Notebook

After the workload has been successfully created, access the environment using the Service URL provided on the workload details page.

You can verify your development environment and perform basic computational tests by creating a new notebook in the default working directory.

1. Verifying the Environment

Execution Code:

import torch
import tensorflow as tf
import transformers
# Check Version print(f"PyTorch: {torch.__version__}")
print(f"TensorFlow: {tf.__version__}")
print(f"Transformers: {transformers.__version__}")
# Check GPU print(f"PyTorch GPU: {torch.cuda.is_available()}")
print(f"TensorFlow GPU: {tf.config.list_physical_devices('GPU')}")

Expected Results:

PyTorch: 2.0.1+cu118
TensorFlow: 2.12.0
Transformers: 4.28.0
PyTorch GPU: True
TensorFlow GPU: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

2. PyTorch Environment Test

Execution Code:

import torch
if torch.cuda.is_available():
    # GPU Information    print(f"GPU: {torch.cuda.get_device_name(0)}")
    print(f"Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.2f} GB")
    # Matrix Operation Test    a = torch.tensor([[1.0, 2.0], [3.0, 4.0]]).cuda()
    b = torch.tensor([[5.0, 6.0], [7.0, 8.0]]).cuda()
    # Matrix Multiplication    c = torch.matmul(a, b)
    print("\nMatrix A:")
    print(a.cpu().numpy())
    print("\nMatrix B:")
    print(b.cpu().numpy())
    print("\nMatrix Multiplication Result (A × B):")
    print(c.cpu().numpy())

Expected Results:

GPU: NVIDIA GeForce RTX 3060
Memory: 12.88 GB

Matrix A:
[[1. 2.]
 [3. 4.]]

Matrix B:
[[5. 6.]
 [7. 8.]]

Matrix Multiplication Result (A × B):
[[19. 22.]
 [43. 50.]]

🔗 View More PyTorch Tutorials

3. TensorFlow Environment Test

Execution Code:

import tensorflow as tf
if tf.config.list_physical_devices('GPU'):
    # GPU Meomory Configuration   gpus = tf.config.experimental.list_physical_devices('GPU')
    tf.config.experimental.set_memory_growth(gpus[0], True)
    # Simple Computation Test    with tf.device('/GPU:0'):
        x = tf.random.normal([1000, 1000])
        y = tf.matmul(x, x)
        print("GPU Computation Test Completed Successfully")

Expected Results:

GPU Computation Test Completed Successfully

🔗 Getting Started with TensorFlow

4. Transformers Environment Test

Execution Code:

from transformers import AutoModel, AutoTokenizer
model_name = "bert-base-uncased"tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)
if torch.cuda.is_available():
    model = model.cuda()
    print("Model successfully loaded onto GPU")

Expected Results:

Model successfully loaded onto GPU

🔗 View More Transformers Tutorials

---

Warning

Please be aware that all data and environmental settings currently in use will not be saved when the workload is terminated.
Training data and results can be stored using a persistent volume (PV).

References & Resources

• PyTorch Documentation
• TensorFlow GPU Guide
• Transformers Documentation

Contact Support

If you encounter any issues or require further assistance, please contact the gcube Support Team 📧 gcube.ai@data-alliance.com