MyLLM: A Complete Open-Source Framework for Building Large Language Models from Scratch

MyLLM is an open-source project for building large language models from scratch, providing a complete pipeline from tokenizer training to RLHF reinforcement learning, helping developers deeply understand every detail of the Transformer architecture.

In the current large language model ecosystem, frameworks like Hugging Face, PyTorch Lightning, and TRL are quite mature, but they encapsulate a lot of low-level details for ease of use. For researchers and developers who want to truly understand the working principles of Transformers, these "black box" abstractions have instead become learning barriers.

The MyLLM project was born out of this need, with its core philosophy being "From Zero to Hero"—allowing users to truly understand the complete technology stack of modern large language models by implementing each component with their own hands. This project is not just a framework, but also a systematic learning path.

MyLLM adopts a layered architecture design, breaking down the complex large model training process into clear and readable modules:

• model.py: Defines the core model structure in GPT/LLaMA style
• api.py: Provides the LLM class, supporting functions like loading, text generation, batch generation, etc.
• Configs/: Uses dataclass to define ModelConfig and GenerationConfig
• Tokenizers/: Supports GPT2, LLaMA2, LLaMA3, and trainable tokenizers
• Train/: Contains training engines like SFT, DPO, PPO
• utils/: Loaders, samplers, weight mappers, and model registries

The training module uses a plug-in design and supports multiple training paradigms:

• SFTTrainer: Supervised Fine-Tuning Trainer (fully implemented)
• DPOTrainer: Direct Preference Optimization (reserved in the framework)
• PPOTrainer: Proximal Policy Optimization/RLHF (reserved in the framework)
• Accelerator: Supports multiple acceleration schemes like single GPU, DDP, DeepSpeed, FSDP

MyLLM provides three progressive learning paths to meet the needs of users at different stages:

Contains 21 carefully designed Jupyter Notebooks covering every step from word embedding to attention mechanisms, and then to complete model training. Each notebook is equipped with detailed theoretical explanations and runnable code examples.

Breaks down complex concepts into independent experimental units, each module focusing on one core concept such as positional encoding, multi-head attention, layer normalization, etc. This "master one concept at a time" design reduces the learning curve.