YAML Metadata
Warning:
empty or missing yaml metadata in repo card
(https://huggingface.co/docs/hub/datasets-cards)
Profiling Result Visualization Dataset
π Introduction
This repository is a part of "MoE_expert_selection_trace Repository". It provides analysis and visualization results of our profiled expert selection trace for MoE LLM on the MMLU dataset, including Llama4-Maverick, DeepSeek-R1, and Qwen3-235B. For a deeper understanding of the analysis, please refer to our paper.
Key Components:
cross_token_heatmap/β Expert selection heatmap across two adjacent tokens. This corresponds to token-level temporal relations in our paper. Results for prefill and decode stages are presented separately.column_by_layer/β Expert selection heatmap across two adjacent layers. This corresponds to layer-level temporal relations in our paper. Results for prefill and decode stages are presented separately.same_layer_heatmap/β Co-activation heatmap for experts. This corresponds to spatial relations for expert pairs in our paper.cross_layer_heatmap/β Activation frequency for different experts, presented as column figures. This corresponds to spatial relations for single experts in our paper.
π Dataset Structure
Top-Level Layout
profiling_result_fig/
βββ meta-llama
β βββ Llama-4-Maverick-17B-128E-Instruct
β
βββ cognitivecomputations
β βββ DeepSeek-R1-AWQ
β βββ cross_token_heatmap
β β βββ mmlu
β β βββ decode
β β β βββ xxx.png
β β β βββ xxx.txt
β β β βββ ...
β β β
β β βββ prefill
β β βββ prefill_decode_corr.txt
β βββ same_layer_heatmap
β βββ cross_layer_heatmap
β βββ column_by_layer
β
βββ Qwen
βββ Qwen3-235B-A22B-FP8
π File Naming and Domains
The subfolders are named after academic or professional domains from the MMLU benchmark and related datasets. Examples:
Heatmap Files:
There are five types of files:
layer_*.pngβ The original heatmap, reflecting the conditional probability of two activated experts.layer_*_avg.pngβ Normalized heatmap with each value divided by the average value of its corresponding column, eliminating vertical white lines caused by frequently selected experts.layer_*_skew.txtβ Accumulated frequency of the most popular expert pairs, calculated by aggregating frequency.layer_*_cnt_skew.txtβ Accumulated frequency of the most popular expert pairs, calculated by aggregating count. Similar tolayer_*_skew.txt, but more accurate.prefill_decode_corr.txtβ Correlation ratio between the prefill stage and decode stage.
Column Figures:
There are three types of files:
layer_*_prefill.pngβ Statistical results for the prefill stage only.layer_*_decode.pngβ Statistical results for the decode stage only.layer_*_both.pngβ Statistical results considering both prefill and decode stages.
π Citation
If you use this dataset in your research or project, please cite it as:
@misc{yu2025orderschaosenhancinglargescale,
title={Orders in Chaos: Enhancing Large-Scale MoE LLM Serving with Data Movement Forecasting},
author={Zhongkai Yu and Yue Guan and Zihao Yu and Chenyang Zhou and Shuyi Pei and Yangwook Kang and Yufei Ding and Po-An Tsai},
year={2025},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2510.05497},
}
- Downloads last month
- 7