Feature Overview:

• Deliver a modular training techniques library supporting SFT, preference tuning, reasoning optimization, and other advanced training methods developed by the Red Hat AI Innovation team.
• Build a training hub python package with all dependencies. This feature aims to package all Training Hub dependencies into a Red Hat-maintained Python Package Index (PyPI). The primary purpose is to enable OpenShift AI customers to directly pull these dependencies to build their own custom container images for AI workloads. This addresses the challenge of ensuring compatibility and simplifying the user experience when customers move beyond pre-built images.

Goals:

• Provide standalone training capabilities
• Support multiple training techniques developed by the Red Hat AI Innovation team
• Enable direct PyTorch primitive access
• Maintain technique modularity and independence

Requirements:

• Pip installable standalone library
• Python SDK with technique-specific APIs
• CLI interface for training operations
• Support for SFT, Async GRPO, DPO, and other techniques developed by the Red Hat AI Innovation team
• Direct PyTorch integration without high-level abstractions

 
 
MVP 3.0:

• Package all dependencies for instructlab.training and continual training (Orthogonal Subspace Learning For Continual Learning Scenarios )
• Training Hub : https://github.com/Red-Hat-AI-Innovation-Team/training_hub 

Dependencies for 3.0:

#ALL DEPS ACROSS training-hub, instructlab-training, mini-trainer

torch>=2.6
typer
deprecated
ninja
numba
packaging>=20.9
wheel>=0.43
pyyaml
py-cpuinfo
transformers>=4.45.2
datasets>=2.15.0
numpy>=1.26.4,<2.0.0
rich
trl>=0.9.4
peft
pydantic>=2.7.0
aiofiles>=23.2.1
accelerate>=0.34.2

1. OUR LIBRARIES (RELYING ON ONLY ABOVE REQUIREMENTS)
   instructlab-training>=0.11.1
   mini-trainer (NO VERSION YET)
2. THE TRAINING HUB PACKAGE ITSELF
   training-hub (NO VERSION YET)
3. FOR CUDA
   bitsandbytes>=0.43.1
   liger-kernel>=0.5.4
   flash-attn>=2.8.2
4. OUR LIBRARIES WITH CUDA EXTRAS (RELYING ON ONLY THE ABOVE EXTRA REQUIREMENTS)
   instructlab-training[cuda]>=0.11.1
   mini-trainer[cuda] (NO VERSION YET)
   training-hub[cuda] (NO VERSION YET)

 training-hub package must install and run successfully on ARM64 (aarch64) CPUs in addition to x86_64.