miles-rl-training
Train large AI models with less memory in miles.
What it does
Guides the configuration and troubleshooting of low-precision and speculative reinforcement learning training with miles.
When to use it
Use for large MoE training, FP8 or INT4 quantization-aware training, exact train-inference alignment, or speculative RL throughput.
How to use it
It changes how the agent configures and runs miles training using model checkpoints and prompt data.
What you provide
- An existing project
- Your files
Uses
What you need · 6
The large MoE training workflow requires H100 or H200 GPUs with FP8 support.
The large MoE workflow requires a Docker environment with miles; the recommended setup uses docker to pull and run its image.
Access · 2
This skill
https://github.com/radixark/miles
Checks miles configuration on GitHub
radixark/miles:latest
Starts the latest miles container