IronCore: A Full-Stack LLM Training Framework for Individual Developers

IronCore is a full-stack LLM training framework built from scratch by individual developers. It supports the complete workflow from pre-training to alignment, covering advanced algorithms like distributed training, tensor parallelism, expert parallelism, DPO, and GRPO—all driven by YAML configurations. The project aims to help developers deeply understand the underlying principles of LLM training and fill the learning gap caused by the high encapsulation level of existing frameworks.

In today's era of rapid LLM technology development, most developers can only call APIs and struggle to understand the underlying principles of training. Existing frameworks like Transformers and DeepSpeed have high encapsulation levels, which are not conducive to learners grasping core concepts such as distributed training, parallel strategies, and alignment algorithms. IronCore draws inspiration from NVIDIA Megatron-LM and HuggingFace Transformers, with the goal of enabling developers to truly understand the internal mechanisms of LLM training by implementing every component themselves.

IronCore provides a complete training pipeline covering multiple stages:

• Training Modes: Pre-training, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), GRPO;
• Data Preprocessing: Supports FIM/PSM formats, flexible tokenization and data splitting;
• Parallel Strategies: Tensor parallelism, expert parallelism, data parallelism, multi-node training, FSDP;
• Model Architectures: GPT-2/3, LLaMA, Gemma, Qwen, Phi, etc.;
• MoE Support: Expert routing, Z-loss regularization, expert parallelism;
• PEFT: LoRA implementation, TP-aware fine-tuning;
• Alignment Algorithms: DPO, GRPO (including KL penalty, multi-epoch replay, etc.), multi-backend reward models;
• Optimizers: Muon optimizer, AdamW hybrid optimization, ZeRO-1;
• Checkpoint Management: Native/distributed checkpoints, HF interoperability, KV caching, MFU monitoring.

The greatest value of IronCore lies in its educational significance:

• Distributed Training Practice: By implementing strategies like TP/EP/DP, understand issues such as all-reduce communication, load balancing, and parallel combination;
• Alignment Algorithm Analysis: The GRPO implementation allows developers to master core challenges like distribution shift handling, IS ratio clipping, and sample efficiency improvement;
• Engineering Closed Loop: Covers the entire workflow from data preprocessing to deployment, learning practices like efficient data loading, stable distributed training, and training efficiency monitoring.

IronCore is suitable for the following scenarios:

1. LLM Researchers: Deeply understand the principles of training algorithms;
2. AI Engineers: Customize training workflows;
3. Learners: Master core concepts like distributed training and alignment technologies;
4. Resource-Constrained Teams: Individuals or small teams training models on limited hardware.

IronCore demonstrates the engineering depth of individual developers supported by modern AI infrastructure. It lowers the entry barrier through Docker containerization, NGC PyTorch images, and detailed configuration documents. For developers who want to advance from "using LLMs" to "understanding LLMs", it is recommended to use IronCore as a learning platform to deeply explore LLM training technologies.