FLUX Schnell H100 Optimization Testing Framework

Overview

This directory contains 8 parallel testing environments for evaluating different FLUX Schnell optimization approaches on H100 GPUs. Each subfolder has its own virtual environment and detailed testing plan.

Testing Structure

01_baseline_torch_compile/ ✅ COMPLETED

Target: Establish baseline with standard torch.compile optimizations

• Status: 100% complete - torch.compile working, memory constraints identified
• Key Features: H100-specific inductor flags, regional compilation
• Virtual Environment: Create fresh when needed: python -m venv venv_torch_compile

02_flux_fast/

Target: Official Hugging Face flux-fast optimizations

• Expected Performance: 2.5x speedup (~4s)
• Key Features: Flash Attention v3, FP8, torch.export + AOTI
• Virtual Environment: Create fresh when needed: python -m venv venv_flux_fast

03_torchao_quantization/

Target: TorchAO quantization techniques

• Expected Performance: 3.4x speedup (~2.966s)
• Key Features: FP8 row-wise quantization, INT8 dynamic
• Virtual Environment: Create fresh when needed: python -m venv venv_torchao

04_xdit_multi_gpu/ 🔄 60% COMPLETE

Target: Multi-GPU parallel inference

• Expected Performance: 2.6x on 4xH100, sub-1s on 8xH100
• Key Features: Ulysses/Ring parallelism, PipeFusion
• Virtual Environment: Create fresh when needed: python -m venv venv_xdit

05_gguf_quantization/

Target: GGUF block-wise quantization

• Expected Performance: Memory efficiency, moderate speedup
• Key Features: Q2_K/Q4_K/Q8_0 quantization, dynamic dequantization
• Virtual Environment: Create fresh when needed: python -m venv venv_gguf

06_svdquant_nunchaku/ ❌ H100 INCOMPATIBLE

Target: Aggressive 4-bit quantization

• Status: Not compatible with H100 GPUs (lacks 4-bit tensor cores)
• Key Features: W4A4 quantization, low-rank branch, LoRA support
• Virtual Environment: Skip - incompatible with hardware

07_diffusers_native/

Target: Built-in diffusers optimizations

• Expected Performance: 20-40% speedup, significant memory reduction
• Key Features: SDPA, xFormers, model offloading, VAE optimizations
• Virtual Environment: Create fresh when needed: python -m venv venv_diffusers_native

08_tensorrt_optimization/

Target: NVIDIA TensorRT engine optimization

• Expected Performance: Hardware-optimized inference (~2-3s)
• Key Features: FP8 quantization, kernel fusion, engine compilation
• Virtual Environment: Create fresh when needed: python -m venv venv_tensorrt

Parallel Testing Instructions

For Each Sub-Agent:

1. Navigate to assigned folder:

   cd flux_optimization_tests/[folder_name]
   

2. Read the README.md for detailed instructions

3. Follow the plan exactly:

   • Check prerequisites
   • Create virtual environment
   • Install dependencies
   • Implement optimizations
   • Run benchmarks
   • Generate reports

Standardized Output Format

Each test should produce:

• Performance metrics (JSON format)
• Sample images for quality comparison
• Memory usage logs
• Error logs (if any)
• Summary report with recommendations

Coordination

• Each test runs independently in its own environment
• Results will be aggregated for final comparison
• Focus on your assigned optimization technique
• Document any issues or unexpected findings

Success Criteria

• All 8 optimization approaches tested
• Performance comparison matrix generated
• Quality assessment completed
• Memory usage analysis documented
• Production recommendations provided