llm-forge-new 0.2.1

Config-driven, YAML-first open-source LLM training platform

llm-forge

Build your own AI model. No ML expertise required.

llm-forge lets you fine-tune large language models by editing a single YAML file and dropping in your dataset. One config file controls everything: data cleaning, training, evaluation, and serving.

Clone and Run (5 Minutes)

No config editing needed. Copy-paste these commands to train your first model:

# 1. Clone and install
git clone https://github.com/Nagavenkatasai7/llm-forge.git
cd llm-forge
pip install -e ".[all]"

# 2. Login to HuggingFace (one-time, for downloading models)
huggingface-cli login

# 3. Train Llama-3.2-1B on Alpaca dataset (~30 min on GPU)
llm-forge train --config configs/demo_lora_llama.yaml --verbose

# 4. Chat with your fine-tuned model
llm-forge serve --config configs/demo_lora_llama.yaml \
  --model-path outputs/demo-lora-llama3.2-1b/merged

Your fine-tuned model appears in outputs/. Chat UI opens at http://localhost:7860.

No GPU? Test on CPU in 5 minutes:

# Uses a tiny 135M model and 10 sample records (included in repo)
llm-forge train --config configs/quickstart_tiny.yaml --verbose

Prerequisites

Before installing, make sure you have:

• Python 3.10-3.12 (python --version)
• NVIDIA GPU with 8+ GB VRAM (nvidia-smi) - or CPU for testing
• Git (git --version)
• 20 GB free disk space
• HuggingFace account (free signup) - needed for gated models like Llama

No GPU? You can still test with mode: "qlora" on CPU (slow but works), or use a free GPU on Google Colab.

Installation

# From source (recommended)
git clone https://github.com/Nagavenkatasai7/llm-forge.git
cd llm-forge
pip install -e ".[all]"

# Or from PyPI
pip install llm-forge[all]

How It Works

You edit this:                  llm-forge handles the rest:
+-----------------+
|   config.yaml   |  ------>   Data Cleaning -> Training -> Evaluation -> Serving
+-----------------+

1. Generate a starter config:

llm-forge init --template lora

2. Edit config.yaml - set your model, data path, and hyperparameters. Every field has inline documentation.

3. Validate before training (catches errors early):

llm-forge validate config.yaml

4. Train:

llm-forge train --config config.yaml

llm-forge auto-detects your hardware and optimizes accordingly.

Example: Train on Your Own Data

Create a JSONL file with instruction-output pairs:

{"instruction": "Summarize this article", "input": "The economy grew...", "output": "Economic growth..."}
{"instruction": "Translate to French", "input": "Hello world", "output": "Bonjour le monde"}

Then create a config (config.yaml):

model:
  name: "meta-llama/Llama-3.2-1B"
  max_seq_length: 2048

data:
  train_path: "./data/train.jsonl"    # Your data file
  format: "alpaca"

training:
  mode: "lora"                        # Memory-efficient fine-tuning
  output_dir: "./outputs/my-model"
  num_epochs: 3
  learning_rate: 2.0e-4

Run llm-forge train --config config.yaml and you're done.

Features

Getting Started

Feature	Description
YAML-First Config	Single config file controls the entire pipeline
Hardware Auto-Detection	Works seamlessly from RTX 3060 to H100 clusters
Smart Validation	Catches config errors with actionable suggestions

Training

Feature	Description
LoRA / QLoRA	Memory-efficient fine-tuning (train 7B models on 24GB GPUs)
Full Fine-Tuning	Unrestricted parameter updates when you have the VRAM
Pre-Training	Train language models from scratch
DPO Alignment	Direct Preference Optimization for RLHF
NEFTune	Noise-based regularization for better generalization

Data & Evaluation

Feature	Description
Data Cleaning	Unicode fixing, deduplication, PII redaction, toxicity filtering
Evaluation	lm-evaluation-harness benchmarks with HTML reports
RAG Pipeline	Chunking, embeddings, hybrid retrieval, reranking

Production

Feature	Description
Distributed Training	FSDP, DeepSpeed ZeRO, Megatron-Core
Serving	Gradio UI, FastAPI REST API, vLLM high-throughput
Model Export	safetensors, GGUF, ONNX formats
HPC Support	SLURM scripts, Singularity containers

Recommended Starting Models

Model	Size	Best For	GPU Needed
HuggingFaceTB/SmolLM2-135M	135M	Quick testing, learning	Any (even CPU)
meta-llama/Llama-3.2-1B	1B	General fine-tuning (recommended)	8+ GB VRAM
Qwen/Qwen2.5-1.5B	1.5B	Multilingual, code	12+ GB VRAM
meta-llama/Llama-3.2-3B	3B	Higher quality results	16+ GB VRAM
microsoft/phi-3-mini-4k-instruct	3.8B	Instruction tuning	24+ GB VRAM

Hardware Compatibility

GPU	VRAM	Recommended Mode	Max Model Size
RTX 3090	24 GB	QLoRA (4-bit)	7B
RTX 4090	24 GB	LoRA / QLoRA	7B
A100 40GB	40 GB	LoRA	13B
A100 80GB	80 GB	Full fine-tune	7B
H100 80GB	80 GB	Full fine-tune + FP8	13B
Multi-GPU	Varies	FSDP / DeepSpeed	70B+
CPU only	N/A	QLoRA (slow)	1B

CLI Reference

llm-forge init [--template lora|qlora|pretrain|rag|full]   # Generate starter config
llm-forge validate config.yaml                              # Validate + hardware check
llm-forge train --config config.yaml [--verbose] [--dry-run]  # Train model
llm-forge eval --config config.yaml                         # Run benchmarks
llm-forge serve --config config.yaml                        # Launch chat UI
llm-forge export --config config.yaml --format gguf         # Export model
llm-forge clean --config config.yaml                        # Data cleaning only
llm-forge rag build --config config.yaml                    # Build RAG index
llm-forge rag query "question" --config config.yaml         # Query RAG
llm-forge info                                              # System + GPU info
llm-forge hardware                                          # Hardware summary

Example Configs

Ready-to-use configs in configs/:

Config	Use Case
quickstart_tiny.yaml	5-min test: SmolLM2-135M on sample data (works on CPU)
demo_lora_llama.yaml	Demo: Llama-3.2-1B on Alpaca (copy-paste ready)
benchmark_smollm_135m.yaml	Benchmark: SmolLM2-135M (<5 min on A100)
benchmark_smollm_360m.yaml	Benchmark: SmolLM2-360M (~10 min on A100)
benchmark_tinyllama_1b.yaml	Benchmark: TinyLlama-1.1B (~15 min on A100)
benchmark_qlora_phi2.yaml	Benchmark: Phi-2 QLoRA 4-bit (~20 min on A100)
benchmark_llama_1b_full.yaml	Benchmark: Llama-3.2-1B full LoRA (~30 min on A100)
example_lora.yaml	LoRA fine-tuning template
example_qlora.yaml	QLoRA for memory-constrained GPUs
example_pretrain.yaml	Pre-training from scratch
example_rag.yaml	RAG pipeline
example_medical_domain.yaml	Medical domain fine-tuning
example_legal_domain.yaml	Legal domain fine-tuning
example_code_domain.yaml	Code generation fine-tuning

Troubleshooting

Problem	Solution
OutOfMemoryError	Use mode: "qlora" in your config
401 Unauthorized from HuggingFace	Run huggingface-cli login
flash_attention_2 not found	OK - llm-forge auto-falls back to SDPA
Training interrupted	Rerun same command - auto-resumes from last checkpoint
BF16 not supported	llm-forge auto-falls back to FP16 on older GPUs
Model download fails	Rerun - downloads auto-resume

Docker

# GPU training
docker build -t llm-forge:gpu -f docker/Dockerfile.gpu .
docker run --gpus all -v $(pwd)/outputs:/app/outputs \
  llm-forge:gpu train --config configs/demo_lora_llama.yaml

# Docker Compose (training + serving + vector DB)
docker compose -f docker/docker-compose.yml up

HPC / SLURM

# Single GPU on Hopper cluster
sbatch scripts/slurm/train_demo.sbatch

# Multi-node distributed
sbatch scripts/slurm/train_multi_node.sbatch

See Hopper Cluster Guide for detailed setup.

Documentation

Doc	For
Quickstart Guide	First-time setup walkthrough
Configuration Reference	All YAML fields explained
Data Preparation	Preparing your dataset
Training Guide	LoRA vs QLoRA vs full fine-tuning
Evaluation Guide	Benchmarking your model
Deployment Guide	Serving in production
Distributed Training	Multi-GPU scaling
API Reference	Python API docs

Contributing

See CONTRIBUTING.md for development setup and guidelines.

Citation

@software{llm_forge,
  author = {Chennu, Naga Venkata Sai},
  title = {llm-forge: Config-Driven LLM Training Platform},
  year = {2026},
  url = {https://github.com/Nagavenkatasai7/llm-forge}
}

License

Apache License 2.0

Author

Naga Venkata Sai Chennu - George Mason University

• GitHub: @Nagavenkatasai7
• Email: [e-mail]