Minimal repository to demonstrate fast LoRA inference with Flux.1-dev using different settings that can help with speed or memory efficiency. Please check the accompanying blog post at this URL.
The included benchmark script allows to experiment with:
- FlashAttention3
torch.compile- Quantization
- LoRA hot-swapping
- CPU offloading
| Option | Time (s) ⬇️ | Speedup (vs baseline) ⬆️ | Notes |
|---|---|---|---|
| baseline | 7.8910 | – | Baseline |
| optimized | 3.5464 | 2.23× | Hotswapping + compilation without recompilation hiccups (FP8 on by default) |
| no_fp8 | 4.3520 | 1.81× | Same as optimized, but with FP8 quantization disabled |
| no_fa3 | 4.3020 | 1.84× | Disable FA3 (flash-attention v3) |
| baseline + compile | 5.0920 | 1.55× | Compilation on, but suffers from intermittent recompilation stalls |
| no_fa3_fp8 | 5.0850 | 1.55× | Disable FA3 and FP8 |
| no_compile_fp8 | 7.5190 | 1.05× | Disable FP8 quantization and compilation |
| no_compile | 10.4340 | 0.76× | Disable compilation: the slowest setting |
The requirements for this repository are listed in the requirements.txt, please ensure they are installed in your Python environment, e.g. by running:
python -m pip install -r requirements.txt.
Optionally, use FlashAttention3 for even better performance. This requires a Hopper GPU (e.g. H100). Follow the install instructions here.
Run the benchmarks using the provided run_benchmark.py script. To check the available arguments, run:
python run_benchmark.py --helpIf you want to run a battery of different settings, some shell scripts are provided to achieve that. Use run_experiments.sh if you have a server GPU like an H100. Use run_exps_rtx_4090.sh if you have a consumer GPU with 24 GB of memory, like an RTX 4090. The benchmark data and sample images are stored by default in the results/ directory.
The inference_lora.py script implements the optimizations in sequence and it is geared towards an H100. Refer to it for a simpler reference than run_benchmark.py. Users should
only refer to this script in case they are not interested in conducting without running
benchmarking.
