A lightweight, AI-native training framework for large language models, StepTronOSS is designed for fast iteration and reproducible experiments. It targets researchers and engineers, offering modular configuration for Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and evaluation workflows, with PyTorch as its sole core dependency.

How It Works

The framework employs a config-driven approach with dynamic validation tools like cfshow and sanity_check for rapid development and inspection. It supports multi-task orchestration and extensible data, optimizer, and model stacks for research agility. Key principles include stateless configurations, runtime state management, and cross-node linkage via Ref. Operator-level replacements enable performance acceleration through custom kernels.

Quick Start & Requirements

Installation requires redis-server and uv (recommended). Core commands involve uv sync install, uv run cfshow for config inspection, uv run torchrun for single-task experiments, and uv run tools/mp_run.py for multi-task setups, which requires STEPTRON_MEET_DIR to be set to a shared directory.

• Launch Guide (EN): docs/LAUNCH_EXPERIMENTS.md
• SFT Data Prep (EN): docs/SFT_DATA_PREPARATION_EN.md

Highlighted Details

• Config-driven experiments with dynamic validation (cfshow, sanity_check).
• Multi-task orchestration and flexible launch tooling.
• Extensible data, optimizer, and model stacks for rapid research iteration.
• Operator-level optimization support (e.g., flash-attn, grouped_gemm).
• AI-native workflow with agent contribution guidance (AGENTS.md).
• Reference configurations for Qwen3 8B and Step3.5 Flash.

Maintenance & Community

No specific community links, contributors, or roadmap details were found in the provided README. The project includes AGENTS.md to guide AI agents in contributing code, suggesting a focus on AI-assisted development.

Licensing & Compatibility

The provided README does not specify a license type or compatibility notes for commercial use.

Limitations & Caveats

The framework is actively under development. Evaluation, RLHF, and Triton kernel implementations are marked as incomplete. Users should note the absence of these features in the current release.