Skip to content

README.md

Latest commit

 

History

History
103 lines (76 loc) · 3.55 KB

README.md

File metadata and controls

103 lines (76 loc) · 3.55 KB
Raw

SFT Finetuner

No-code Supervised Fine-Tuning tool designed for NVIDIA DGX Spark / Asus Ascent GX10 (GB10 · 128 GB Unified RAM)

An interactive CLI tool that guides you through the entire SFT workflow — from model selection to training to uploading your fine-tuned model — without writing a single line of code.

Features

  • Interactive CLI — step-by-step prompts, no coding required
  • LoRA / QLoRA / Full Fine-Tuning — choose the strategy that fits your memory budget
  • Safetensor Validation — automatically rejects GGUF and MLX models
  • Flexible Dataset Support — conversational (messages), prompt-completion, plain text, and legacy System/User/Assistant formats
  • Local & HuggingFace Datasets — load from disk (csv, json, jsonl, parquet) or directly from the Hub
  • Automatic Format Conversion — legacy datasets are converted to the modern conversational format on the fly
  • BF16 & Sequence Packing — optimized for Blackwell architecture throughput
  • LoRA Merge & Hub Upload — merge adapter weights and push the final model to HuggingFace Hub in one flow

Hardware Requirements

Component Specification
Device NVIDIA DGX Spark or Asus Ascent GX10
GPU NVIDIA GB10 Grace Blackwell Superchip
Memory 128 GB Unified RAM
CUDA 13.0 (via PyTorch nightly)
Python 3.10+

The tool can also run on other CUDA-capable GPUs, but optimal performance and memory defaults are tuned for the GB10 128 GB unified memory architecture.

Quick Start

git clone https://github.com/alicankiraz1/DGX-Spark-Asus-Ascent-Nvidia-GB10-SFT-Finetuner.git
cd DGX-Spark-Asus-Ascent-Nvidia-GB10-SFT-Finetuner

chmod +x setup_and_run.sh
./setup_and_run.sh

The setup script will:

  1. Create a Python virtual environment
  2. Install PyTorch nightly with CUDA 13.0 support
  3. Install all project dependencies
  4. Verify GPU and CUDA availability
  5. Launch the SFT Finetuner

Supported Models

Any HuggingFace model that ships with safetensor (.safetensors) or legacy .bin weight files. Examples:

  • meta-llama/Llama-3.3-70B-Instruct
  • mistralai/Mistral-7B-Instruct-v0.3
  • google/gemma-3-9b-it
  • Qwen/Qwen3.5-9B

Models in GGUF or MLX format are automatically detected and rejected.

Supported Dataset Formats

Format Required Columns
Conversational messages (list of {role, content} dicts)
Prompt-Completion prompt, completion
Language Modeling text
Legacy (LLMRipper) System, User, Assistant

Fine-Tuning Strategies

Strategy Description Memory Usage
LoRA Low-Rank Adaptation with rsLoRA — recommended default Medium
QLoRA 4-bit NF4 quantization + LoRA — lowest memory footprint Low
Full All parameters updated — only for smaller models High

Project Structure

├── SFTFinetuner.py      # Main interactive fine-tuning tool
├── setup_and_run.sh     # Environment setup & launcher script
├── requirements.txt     # Python dependencies
└── README.md            # This file

Dependencies

  • PyTorch (nightly, CUDA 13.0)
  • Transformers
  • TRL (Transformer Reinforcement Learning)
  • PEFT (Parameter-Efficient Fine-Tuning)
  • Datasets
  • BitsAndBytes
  • Accelerate
  • HuggingFace Hub
  • Safetensors
  • Pyfiglet

Author

Alican Kiraz

License

This project is licensed under the MIT License.