File size: 3,629 Bytes
74cbb3d b4999b8 e0eed95 74cbb3d 4fe7f02 74cbb3d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 |
---
license: mit
tags:
- benchmark
- systems-ml
- distributed-training
- muon
- optimizer
- performance-analysis
---
# 🔬 Distributed Muon: Field Notes & Reproducibility Artifacts
**Code, Performance Traces, and Analysis Logs**
This repository contains the raw engineering artifacts for the deep-dive investigation: **"Reproducing and Validating Distributed Muon"**.
It serves as the **proof of work** for the performance claims regarding the Muon optimizer's communication efficiency and computational overhead in a distributed setting (Data Parallel + Tensor Parallel).
📄 **Read the Full Report:** [Reproducing and Validating Distributed Muon 🐢✨: A Practical Verification of Communication Efficiency Claims](https://medium.com/@jenwei0312/reproducing-and-validating-distributed-muon-a-practical-verification-of-communication-0be4d1d9b893)
🛠️ **Get the Tutorial Code:** [bird-of-paradise/muon-distributed](https://huggingface.co/datasets/bird-of-paradise/muon-distributed)
---
## 📂 Repository Structure
* **`traces/`**: Raw Chrome Trace (`.json`) files generated by PyTorch Profiler. You can load these into `chrome://tracing` or [ui.perfetto.dev](https://ui.perfetto.dev) to visualize the exact CPU/GPU execution timeline.
* `comparison/`: Side-by-side traces of AdamW vs. Muon (Hybrid DP=2/TP=2).
* `distributed_muon/`: Scaling traces for DP=4, TP=4, and Hybrid configurations.
* **`analysis_scripts/`**: The exact Python scripts used to generate the traces and parse the performance metrics.
* **`figures/`**: High-resolution charts and trace visualizations used in the report.
* **`report/`**: A PDF archive of the full technical investigation.
---
## 🔍 Key Findings (Verified in Traces)
The traces in this repository provide empirical evidence for the following:
1. **Communication Efficiency:** Muon (Hybrid DP2/TP2) demonstrates **0.57x** the communication overhead of AdamW on a bandwidth-constrained cluster (PCIe Gen4 x4).
* *Evidence:* Compare `traces/comparison/adamw_fullstep_rank0.json` vs `muon_fullstep_dp2_tp2_rank0.json`.
2. **Optimizer Latency:** The Muon step accounts for **~1.1%** of total training time, validating the paper's "negligible overhead" claim.
3. **Hybrid Scaling:** The `DP=2, TP=2` configuration outperforms pure DP or pure TP on 4 GPUs, balancing memory bandwidth with communication overhead.
---
## 🛠️ How to Reproduce
To run these benchmarks yourself on a 4-GPU cluster:
1. Clone this repository.
2. Install dependencies: `torch`.
3. Run the benchmark script:
```bash
# This will generate new JSON traces in your local directory
python analysis_scripts/muon_vs_adam.py
```
4. Run the performance analysis on included trace files
```bash
python analysis_scripts/performance_comparison.py
```
---
## 🙏 Acknowledgments
- [Mahdi Chaker](https://github.com/mchaker) for generously providing GPU cluster access
- MoonShot AI team for open-sourcing their [PoC implementation](https://github.com/NVIDIA/Megatron-LM/pull/1428/commits/f432fbe45c169aeb5a0805ff6f41e13f989c6730#diff-61c8e9370cb7fd634a4019472368c487898093f5d330375524c76eac15c7390c)
---
📖 Citation
If you use these traces or analysis in your work, please cite:
@misc{wei2025muoneproducibility,
author = {Wei, Jen},
title = {Distributed Muon: Performance Artifacts and Benchmarks},
year = {2025},
publisher = {Hugging Face},
journal = {Hugging Face Datasets},
howpublished = {\url{[https://huggingface.co/datasets/bird-of-paradise/muon-distributed-reproducibility](https://huggingface.co/datasets/bird-of-paradise/muon-distributed-reproducibility)}}
}
|