README.md

This repository contains a lightweight, custom implementation of MuParameterization as described here: https://www.microsoft.com/en-us/research/blog/%C2%B5transfer-a-technique-for-hyperparameter-tuning-of-enormous-neural-networks

$\mu$-Parameterization is a reparameterization of some of the hyperparameters in a Neural Network that allows optimizing hyperparameters on a small model and transferring these optimimal hyperameters to larger models

Hyper-parameters that can be transferred:

• Learning Rate (and Schedule)
• Initialization
• Optimizer params (eg Adam eps/beta)

Dimensions that can be varied:

• Width
• Depth

** The small model must be trained using the same batch size and total number of steps as the model you wish to transfer hyperparameters to **

This implementation is compatible with DDP and FSDP.

Example Usage:

get mup_multipliers:

mup_multipliers = get_mup_multipliers(base_model, model)

use them to scale initial variance

mup_init(model, mup_multipliers)

We build optimizer groups based off the mup_multipliers so that each group uses the correct scaled lr

optimizer_cls = torch.optim.AdamW # can use any optimizer optimizer_param_groups = build_optimizer_param_groups(dist_model, mup_multipliers, **optimizer_kwargs) optimizer = optimizer_cls(optimizer_param_groups)