Nanoflash

Minimal training framework driven by a single YAML config. Uses OmegaConf, Hydra instantiate, and a simple train pipeline. Supports Qwen/Qwen3-1.7B SFT on Nvidia GPU.

Quick Start

# From project root (ensure PYTHONPATH includes project root)
python -m train --config nanoflash/config/qwen3_0.6b.yaml

# Override config via CLI
python -m train --config nanoflash/config/qwen3_0.6b.yaml train.batch_size=8 train.max_steps=500

Requirements

• PyTorch (with CUDA for GPU)
• transformers
• datasets
• omegaconf
• hydra-core

pip install torch transformers datasets omegaconf hydra-core

Config

All components are configured via YAML with class_path pointing to the class or factory:

• model: Causal LM (e.g. Qwen3)
• data: Dataset (e.g. Alpaca-style)
• optimizer, lr_scheduler, loss, collate_fn
• checkpoint, logging

Use ${key.subkey} for interpolation (e.g. ${model.model_name_or_path}).

Project Path

To load components from an external directory, set project in config:

project: /path/to/your/code

This prepends the path to sys.path so class_path can import from that tree.

Adding Components

• New loss/dataset: Implement the class or factory, then set class_path in YAML. No code change in nanoflash.
• New recipe: Implement a class with __init__, setup, train, cleanup; register in nanoflash.pipeline.recipe.RECIPES; set recipe: your_name in YAML.

See design.md for full design. Usage for usage guide.