[EPIC] Setting up a workspace to fine-tune a Gen AI model

[ederign]

User Story

As an AI/ML practitioner, Bella needs to quickly set up an interactive environment to fine-tune a Gen AI model. She browses the list of available WorkspaceKinds, filtering by IDE type and supported frameworks (e.g., JupyterLab, Pytorch).

She then selects a pre-built image with key libraries like CUDA, Hugging Face, and InstructLab tools.

After selecting a suitable image, she chooses a pod configuration with enough GPU, CPU, and memory resources to support model fine-tuning.

During the setup, Bella creates a new secret to securely store her access tokens and mounts a persistent volume to hold training datasets and model outputs.

She launches her environment with just a few clicks, ready to fine-tune her first Gen AI model.

Features

1. [FEATURE] Workspace creation base flow #282

2. [FEATURE] Workspace Kind Selection #283

3. [FEATURE] Workspace Image Selection #284

4. [FEATURE] Workspace Pod Template selection #285

5. [FEATURE] Create and mount a secret inline during a workspace creation flow #286

6. [FEATURE] Create and mount a volume inline during a workspace creation flow #287

[ederign]

/area v2
/area frontend
/area backend

[ederign]

/area controller

[jiridanek]

Bella needs to quickly set up an interactive environment to fine-tune a Gen AI model.

Do we have an example of the actual finetuning itself? Can we assume it's exactly like in https://pytorch.org/torchtune/0.4/tutorials/first_finetune_tutorial.html?

Do we assume some kind of distributed workloads backend running on the cluster that the pytorch connects to.

[ederign]

@jiridanek, for now, let's assume that could be any 'fine-tune' job, i.e., the one that you mention or InstructLab. We can assume that everything is already set up in the cluster. Our focus is more on 'any gen ai workload that requires gpu' (and by consequence, notebooks 2.0 would be a good fit to handle with multiple workspaces running those tasks)