pkgs/MagiCompiler/docs/Wan2.2Benchmark.md

Wan2.2 Benchmark

Executive Summary

This report presents a comprehensive performance evaluation of the Athena framework compared to the baseline LightX2V framework. The benchmarks were conducted using the Wan2.2-TI2V-5B model on NVIDIA H100 hardware.

---

🎯Test Environment & Versioning

Hardware & Settings

Parameter	Value
Hardware	NVIDIA H100
Model	Wan2.2-TI2V-5B
Precision	torch.bfloat16
Inference Steps	50
Resolution	704 × 1280(720p)
FPS	24
CFG	Enabled

Software Versioning

To ensure reproducibility, the following specific commits were used for this benchmark:

Framework	Branch / Tag	Commit
Athena	main	f676ae6
LightX2V	main	33f0f67

🏆 Performance Benchmarks

📊 We compared the iteration speed (seconds per iteration) between Athena and LightX2V across three distinct Context Parallel (CP) configurations.

Configuration	Frames	LightX2V (s/it)	Athena (s/it)	Speedup
CP1	121	1.928	1.69	1.14x 🚀
CP2	121	1.197	1.06	1.13x 🚀
CP4	241	1.767	1.32	1.34x 🚀
CP8	241	1.507	1.35	1.12x 🚀

---

💡 Reproduction Guide

To reproduce the results presented in this report, follow the steps below using the specified commit hashes.

Setup

git clone https://github.com/world-sim-dev/athena
cd athena
git checkout f676ae6
pip install -r requirements.txt

# Clone and install LightX2V (for baseline comparison)
git clone https://github.com/ModelTC/LightX2V
cd lightx2v
git checkout 33f0f67
pip install -r requirements.txt

Running Benchmarks

For Athena, run:

bash ./scripts/run_wan2_2_ti2v_i2v.sh

For LightX2V: Clone the scripts from Benchmark for LightX2V and run:

git clone https://gist.github.com/wtr0504/629388f17ed38d1c12d5ef5c25a15197
bash run_wan.sh

🔎 MagiCompiler Optimization Methodology

Whole Graph Compilation Constant Folding & Dead Code Elimination: Streamlining the computation graph prior to execution. Coarse-grained Kernel Fusion MagiCompiler aggregates multiple smaller operators into larger, fused kernels. This optimization is critical for efficient execution on the GPU. All to All Communication MagiCompiler Uses all_to_all_single (1 communication op per sync point) while LightX2V Uses all_to_all x 3 (3 separate communication ops).