Skip to content

Latest commit

 

History

History
executable file
·
124 lines (69 loc) · 8.33 KB

README.md

File metadata and controls

executable file
·
124 lines (69 loc) · 8.33 KB
Raw

Datalayer

Become a Sponsor

Ξ Datalayer Examples

This repository contains Jupyter notebook examples showcasing scenarios where Datalayer proves highly beneficial. Datalayer allows you to scale Jupyter Kernels from your local JupyterLab or CLI to the cloud, providing the capability to run your code on powerful GPU(s) and CPU(s). 🚀

The Technical validation section delves into system checks and performance benchmarks to ensure optimal GPU and CPU utilization, while the Use cases section explores typical AI scenarios where scaling proves essential.

💡 Note that you can use any notebook within Datalayer without requiring any code changes.

Getting started

pip install datalayer jupyterlab
git clone https://github.com/datalayer/examples.git datalayer-examples
cd datalayer-examples
jupyter lab

Read the documentation website to know more about how setup Datalayer. Don't worry, it is easy 👍
You just need to install the package, open JupyterLab, click on the Jupyter kernels tile in the JupyterLab launcher, create an account, wait a bit for your Kernels to be ready, and then just assign a Remote Kernel from any Notebook kernel picker.

Notebook remote execution

Technical validation

  1. GPU sanity checks
  2. Performance comparison of CPU and GPU serial and parallel execution

1. GPU sanity checks

This notebook contains scripts and tests to perform GPU sanity checks using PyTorch and CUDA. The primary goal of these checks is to ensure that the GPU resources meet the expected requirements.

2. Performance comparison of CPU and GPU serial and parallel execution

This notebook explores the performance differences between serial and parallel execution on CPU and GPU using PyTorch. We'll compare the execution times of intensive computational tasks performed sequentially on CPU and GPU, as well as in parallel configurations.

Use cases

  1. Face detection on YouTube video with OpenCV
  2. Image classification model training with fast.ai
  3. 'Personalized' text-to-image model creation with Dreambooth
  4. Text generation using the Transformers library
  5. Instruction tuning for Mistral 7B on Alpaca dataset

1. OpenCV Face Detection

This example utilizes OpenCV for detecting faces in YouTube videos. It uses a traditional Haar Cascade model, which may have limitations in accuracy compared to modern deep learning-based models. It also utilizes parallel computing across multiple CPUs to accelerate face detection and video processing tasks, optimizing performance and efficiency. Datalayer further enhances this capability by enabling seamless scaling across multiple CPUs.

2. Image Classifier with Fast.ai

This example demonstrates how to build a model that distinguishes cats from dogs in pictures using the fast.ai library. Due to the computational demands of training a model, a GPU is required.

3. Dreambooth

This example uses the Dreambooth method which takes as input a few images (typically 3-5 images suffice) of a subject (e.g., a specific dog) and the corresponding class name (e.g. "dog"), and returns a fine-tuned/'personalized' text-to-image model (source: Dreambooth). To do this fune-tuning process, GPU is required.

4. Text Generation with Transformers

Those notebook examples demonstrate how to leverage Datalayer's GPU kernels to accelerate text generation using Gemma model and the HuggingFace Transformers library.

Transformers Text Generation

This notebook uses Gemma-7b and Gemma-7b-it which is the instruct fine-tuned version of Gemma-7b.

Sentiment Analysis with Gemma

This example demonstrates how you can leverage Datalayer's Cell Kernels feature on JupyterLab to offload specific tasks, such as sentiment analysis, to a remote GPU while keeping the rest of your code running locally. By selectively using remote resources, you can optimize both performance and cost. This hybrid approach is perfect for tasks like sentiment analysis via llm where some parts of the code require more computational resources than others. For a detailed explanation and step-by-step guide on using Cell Kernels, check out our blog post on this specific example.

5. Mistral Instruction Tuning

Mistral 7B is a large language model (LLM) that contains 7.3 billion parameters and is one of the most powerful models for its size. However, this base model is not instruction-tuned, meaning it may struggle to follow instructions and perform specific tasks. By fine-tuning Mistral 7B on the Alpaca dataset using torchtune, the model will significantly improve its capabilities to perform tasks such as conversation and answering questions accurately. Due to the computational demands of fine-tuning a model, a GPU is required.

Datalayer Advanced Features

CLI Execution

Datalayer supports the remote execution of code using the CLI. Refer to this page for more information.

CLI Remote Execution CLI remote execution
Sharing State between Notebook and CLI Remote Notebook Execution

When using the same Kernel, variables defined in a notebook can be used in the CLI and vice versa. This holds also true when using multiple notebooks connected to the same kernel, for example.

Cell Kernel

Datalayer offers the possibility to use cell-specific Kernels, allowing you to execute specific cells with different kernels. This feature optimizes costs by enabling you to, for example, leverage the local CPU for data preparation and reserving the powerful (and often more expensive) GPU resources for intensive computations.

Cell Kernel execution Cell Kernel Execution

The remote GPU Kernel is utilized only for the duration of the cell computation, minimizing costs.