Meta enhancement ideas
Goal
Collect future enhancement ideas that will be branched off into separate "Support XYZ" issues when matured sufficiently.
Future enhancement ideas
Please provide new ideas and updates as comments.
Support for reinforcement learning
Reinforcement learning is used heavily for post-training but appears increasingly also as part of pre-training LLMs. Integrating this into maester would likely require a separate training loop.
Support for quantization-aware training
8-bit, 4-bit, and 1.58-bit quantization-aware training reduce memory consumption and boos performance significantly on modern GPU architectures supporting 8/4 bit operations natively.
Support for DeToNATION
Maester currently supports tensor parallelism and FSDP. There are many promising ideas regarding efficient distributed training that might be integrated via an integration of DeToNATION or porting of important parts into maester.
Current enhancement ideas.
Please provide updates as comments.
Support for distillation
Distillation is an important ingredient of state-of-the-art LLM training pipelines. There are many variants and design choices. The current enhancement issue Support distillation #24 starts the discussion.
Support other optimizers
There is a current enhancement issue Support Muon optimizer #18, which suggests integrating Muon support into maester. Other interesting optimizers might be considered.
Support other model architectures
In addition to the Llama and Gemma 3 architectures, there is a current enhancement PR DeepSeek #30, which enables DeepSeek-style MoE training.
Meta enhancement ideas
Goal
Collect future enhancement ideas that will be branched off into separate "Support XYZ" issues when matured sufficiently.
Future enhancement ideas
Please provide new ideas and updates as comments.
Support for reinforcement learning
Reinforcement learning is used heavily for post-training but appears increasingly also as part of pre-training LLMs. Integrating this into maester would likely require a separate training loop.
Support for quantization-aware training
8-bit, 4-bit, and 1.58-bit quantization-aware training reduce memory consumption and boos performance significantly on modern GPU architectures supporting 8/4 bit operations natively.
Support for DeToNATION
Maester currently supports tensor parallelism and FSDP. There are many promising ideas regarding efficient distributed training that might be integrated via an integration of DeToNATION or porting of important parts into maester.
Current enhancement ideas.
Please provide updates as comments.
Support for distillation
Distillation is an important ingredient of state-of-the-art LLM training pipelines. There are many variants and design choices. The current enhancement issue Support distillation #24 starts the discussion.
Support other optimizers
There is a current enhancement issue Support Muon optimizer #18, which suggests integrating Muon support into maester. Other interesting optimizers might be considered.
Support other model architectures
In addition to the Llama and Gemma 3 architectures, there is a current enhancement PR DeepSeek #30, which enables DeepSeek-style MoE training.