Training Hub

Red Hat's unified library for LLM post-training: SFT, LoRA, and OSFT (continual learning).

Quick Reference

| Task | Command | |------|---------| | Recommend config | python scripts/recommend_config.py --model <model> --hardware <hw> | | Estimate memory | python scripts/estimate_memory.py --model <model> --method sft --hardware h100 | | Validate dataset | python scripts/validate_dataset.py data.jsonl | | Full fine-tuning | from training_hub import sft | | LoRA training | from training_hub import lora_sft | | OSFT (continual) | from training_hub import osft |

Installation

pip install training-hub              # Basic
pip install training-hub[lora]        # LoRA with Unsloth (2x faster)
pip install training-hub[cuda] --no-build-isolation  # CUDA support

Get Started Fast

# Get optimal config for your hardware
python scripts/recommend_config.py \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --hardware rtx-5090

Data Format

Training data must be JSONL with message structure:

{"messages": [{"role": "user", "content": "Q"}, {"role": "assistant", "content": "A"}]}

Validate before training:

python scripts/validate_dataset.py ./training_data.jsonl

For data preparation details, see DATA-FORMATS.md.

Training Methods

Supervised Fine-Tuning (SFT)

Full-parameter fine-tuning. Requires significant VRAM.

from training_hub import sft

result = sft(
    model_path="Qwen/Qwen2.5-7B-Instruct",
    data_path="./training_data.jsonl",
    ckpt_output_dir="./checkpoints",
    num_epochs=3,
    effective_batch_size=8,
    learning_rate=2e-5,
    max_seq_len=2048,
    max_tokens_per_gpu=45000,
)

LoRA Fine-Tuning

Memory-efficient adaptation (up to 2x faster, 70% less VRAM):

from training_hub import lora_sft

result = lora_sft(
    model_path="Qwen/Qwen2.5-7B-Instruct",
    data_path="./training_data.jsonl",
    ckpt_output_dir="./outputs",
    lora_r=16,
    lora_alpha=32,
    num_epochs=3,
    learning_rate=2e-4,
)

QLoRA (4-bit): Add load_in_4bit=True for large models on limited VRAM.

OSFT (Continual Learning)

Adapt without catastrophic forgetting:

from training_hub import osft

result = osft(
    model_path="meta-llama/Llama-3.1-8B-Instruct",
    data_path="./domain_data.jsonl",
    ckpt_output_dir="./checkpoints",
    unfreeze_rank_ratio=0.25,
    effective_batch_size=16,
    learning_rate=2e-5,
)

For all parameters, see ALGORITHMS.md.

Hardware Support

| Hardware | VRAM | Best For | |----------|------|----------| | RTX 5090 | 32GB | 8B LoRA, 70B QLoRA | | DGX Spark | 128GB | 70B SFT | | H100 | 80GB | 14B SFT, 70B LoRA | | 8×H100 | 640GB | 70B SFT |

# Check if your config fits
python scripts/estimate_memory.py \
  --model meta-llama/Llama-3.1-70B-Instruct \
  --method lora \
  --hardware h100 \
  --num-gpus 8

For hardware-specific configs, see HARDWARE.md.

Scaling

Multi-GPU:

result = sft(..., nproc_per_node=8)

Multi-node:

result = sft(..., nnodes=2, node_rank=0, nproc_per_node=8, rdzv_endpoint="0.0.0.0:29500")

For Slurm, Kubernetes, and datacenter deployments, see SCALE.md.

Algorithm Selection

| Scenario | Method | |----------|--------| | First-time fine-tuning, large dataset | SFT | | Memory constrained | LoRA | | Very large model (70B+), limited VRAM | LoRA + QLoRA | | Preserve existing capabilities | OSFT | | Domain adaptation, small dataset | OSFT |

Documentation

| Topic | File | |-------|------| | Hardware profiles & configs | HARDWARE.md | | All algorithm parameters | ALGORITHMS.md | | Data formats & conversion | DATA-FORMATS.md | | Datacenter & cluster setup | SCALE.md | | Model evaluation | EVALUATION.md | | vLLM inference & serving | INFERENCE.md | | Advanced techniques | ADVANCED.md | | Model-specific configs | MODELS.md | | Troubleshooting | TROUBLESHOOTING.md | | Distributed training | DISTRIBUTED.md |

Utility Scripts

| Script | Purpose | |--------|---------| | recommend_config.py | Generate optimal config for model + hardware | | estimate_memory.py | Estimate GPU memory requirements | | validate_dataset.py | Validate JSONL dataset format | | convert_to_jsonl.py | Convert CSV, Alpaca, ShareGPT to JSONL |

Troubleshooting

CUDA OOM: Reduce max_tokens_per_gpu, use LoRA + QLoRA, or add GPUs

Dataset errors: Run python scripts/validate_dataset.py first

LoRA multi-GPU: Requires torchrun --nproc-per-node=N script.py

Training diverges: Lower learning_rate (try 1e-5 for SFT, 1e-4 for LoRA)

For more, see TROUBLESHOOTING.md.