KEP-2401: Kubeflow LLM Trainer V2

[Electronic-Waste]

This is the tracking issue for the Kubeflow LLM Trainer V2, a submodule of Kubeflow Training V2: #2170

We aim to solve:

• KEP-2170: Design Trainer for the LLM Runtimes #2321
• KEP-2170: Create LLM training runtime for Llama 3.2 8B #2212

However, the LLM Trainer V2 design is very complex and needs further discussion. So we decided to open a separate issue tracking it.

• Create the KEP: KEP-2401: Kubeflow LLM Trainer V2 #2410
• KEP-2401: Refactor current train() API #2503
• KEP-2401: Add TorchTuneConfig to train() API #2504
• KEP-2401: Support LoRA/QLoRA/DoRA fine-tuning in LLM Trainer V2 #2505
• KEP-2401: Support mutating dataset preprocessing config in SDK #2506
• KEP-2401: Complement torch plugin to support torchtune config mutation #2507
• KEP-2401: Validate fine-tuning configurations in torch plugin #2508
• KEP-2401: Create LLM Training Runtimes for Llama 3.1 model family #2509
• KEP-2401: Create LLM Training Runtimes for Llama 3.2 model family #2510
• KEP-2401: Create LLM Training Runtimes for Llama 3.3 model family #2591
• KEP-2401: Create torchtune trainer image #2511
• KEP-2401: Determine the tag for torchtune trainer & Add support for multiple accelerators #2518
• KEP-2401: Revisit PVC claim in torchtune CTRs when stateful jobset is ready #2630
• KEP-2401: Revisit DependsOn API in CTRs When Supporting Multiple Ancestor #2592
• KEP-2401: Support loading local LLMs #2641
• Create model exporter for checkpointing and training output #2245

Examples & User Documentation

• KEP-2401: Add Notebook examples for LLM Trainer V2 #2676
• trainer: Add user guide for BuiltinTrainer(TorchTune LLM Trainer) website#4146

KEP Updates:

• Runtime API: feat(doc): add Runtime API design in KEP-2401. #2501
• Train API: fix(doc): Update train() API in KEP-2401 #2536

Jobset Improvements:

• Support multiple items in DependsOn API kubernetes-sigs/jobset#874

Initial Design (Google Doc): Kubeflow Training V2 LLM Trainer Design

/area runtime
/cc @kubeflow/wg-training-leads @deepanker13 @saileshd1402 @seanlaii @helenxie-bit @astefanutti @varshaprasad96 @franciscojavierarceo @thesuperzapper @rimolive @juliusvonkohout @jbottum @varodrig @Doris-xm @truc0

[Electronic-Waste]

/kind feature

[andreyvelich]

@Electronic-Waste Thank you for creating dedicated issue for each task!
I've created new label for you, feel free to use it:

/area llm

[Electronic-Waste]

@andreyvelich Thank you for this!

I'm also wondering if we could pin this issue to let more people track our progress. Currently, this issue has been postponed to the second page of issues. It might be inconvenient for other folks to notice this issue:)

[andreyvelich]

Yes, let me pin it.