Optimizer

Optimizer is a python package that provides:

• PyTorch implementation of recent optimizer algorithms
• with support for parallelism techniques for efficient large-scale training.

Currently implemented

• Parallel Muon with N-D sharding
  • arxiv URL: (TBW)
  • Supports general N-D sharding configurations
    • The implementation is not tied to any specific parallel strategy.
    • Verified from basic FSDP2 setups up to hybrid configurations such as (2 TP + 2 DP-Replicate + 2 DP-Shard).
    • Verified configurations can be found in test_muon.py

Usage

import torch
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from kernels import get_kernel

optimizer = get_kernel("motif-technologies/optimizer")
get_default_muon_param_groups = optimizer.muon.get_default_muon_param_groups

model = None # your model here
fsdp_model = FSDP(model)

# muon, in nature, cannot use 1-d tensor
# we provide helper function to group such tensors
# you can use your own function, if necessary
params = get_default_muon_param_groups(model) # user can write own is_muon_func, if necessary

optim = optimizer.Muon(
    params,
    lr=0.01,
    momentum=0.9,
    weight_decay=1e-4,
)

Test

• Check test/README.md for how to run the tests.

Pre-commit Hooks

This project uses pre-commit to automatically check and format code before commits.

Setup

1. Install pre-commit:

   pip install pre-commit
   

2. Install the git hooks:

   pre-commit install

Once installed, the configured hooks will run automatically on each commit.

Included Hooks

The following tools are run via pre-commit:

• yapf – Python code formatter
• typos – Spell checker for common typos
• isort – Organizes and sorts Python imports
• clang-format – Formats C++/CUDA code (--style=file)
• pymarkdown – Lints and auto-fixes Markdown files
• actionlint – Validates GitHub Actions workflows

Usage

• Run all checks on the entire codebase:

  pre-commit run --all-files
  

• Run a specific hook (example: isort):

  pre-commit run isort --all-files

Test

• There is a simple unittest for Parallel Muon