Skip to content

Optimizations

catboxanon edited this page Jul 19, 2023 · 13 revisions

A number of optimization can be enabled by commandline arguments:

commandline argument explanation
--opt-sdp-attention May results in faster speeds than using xFormers on some systems but requires more VRAM. (non-deterministic)
--opt-sdp-no-mem-attention May results in faster speeds than using xFormers on some systems but requires more VRAM. (deterministic, slightly slower than --opt-sdp-attention and uses more VRAM)
--xformers Use xFormers library. Great improvement to memory consumption and speed. Nvidia GPUs only. (non-deterministic)
--force-enable-xformers Enables xFormers regardless of whether the program thinks you can run it or not. Do not report bugs you get running this.
--opt-split-attention Cross attention layer optimization significantly reducing memory use for almost no cost (some report improved performance with it). Black magic.
On by default for torch.cuda, which includes both NVidia and AMD cards.
--disable-opt-split-attention Disables the optimization above.
--opt-sub-quad-attention Sub-quadratic attention, a memory efficient Cross Attention layer optimization that can significantly reduce required memory, sometimes at a slight performance cost. Recommended if getting poor performance or failed generations with a hardware/software configuration that xFormers doesn't work for. On macOS, this will also allow for generation of larger images.
--opt-split-attention-v1 Uses an older version of the optimization above that is not as memory hungry (it will use less VRAM, but will be more limiting in the maximum size of pictures you can make).
--medvram Makes the Stable Diffusion model consume less VRAM by splitting it into three parts - cond (for transforming text into numerical representation), first_stage (for converting a picture into latent space and back), and unet (for actual denoising of latent space) and making it so that only one is in VRAM at all times, sending others to CPU RAM. Lowers performance, but only by a bit - except if live previews are enabled.
--lowvram An even more thorough optimization of the above, splitting unet into many modules, and only one module is kept in VRAM. Devastating for performance.
*do-not-batch-cond-uncond Prevents batching of positive and negative prompts during sampling, which essentially lets you run at 0.5 batch size, saving a lot of memory. Decreases performance. Not a command line option, but an optimization implicitly enabled by using --medvram or --lowvram.
--always-batch-cond-uncond Disables the optimization above. Only makes sense together with --medvram or --lowvram
--opt-channelslast Changes torch memory type for stable diffusion to channels last. Effects not closely studied.
--upcast-sampling For Nvidia and AMD cards normally forced to run with --no-half, should improve generation speed.

As of version 1.3.0, Cross attention optimization can be selected under settings. xFormers still needs to enabled via COMMANDLINE_ARGS. 2023-06-21 22_53_54_877 chrome

Extra tips (Windows):

Memory & Performance Impact of Optimizers and Flags

This is an example test using specific hardware and configuration, your mileage may vary
Tested using nVidia RTX3060 and CUDA 11.7

Cross-attention Peak Memory at Batch size 1/2/4/8/16 Initial It/s Peak It/s Note
None 4.1 / 6.2 / OOM / OOM / OOM 4.2 4.6 slow and early out-of-memory
v1 2.8 / 2.8 / 2.8 / 3.1 / 4.1 4.1 4.7 slow but lowest memory usage and does not require sometimes problematic xFormers
InvokeAI 3.1 / 4.2 / 6.3 / 6.6 / 7.0 5.5 6.6 almost identical to default optimizer
Doggetx 3.1 / 4.2 / 6.3 / 6.6 / 7.1 5.4 6.6 default
Doggetx 2.2 / 2.7 / 3.8 / 5.9 / 6.2 4.1 6.3 using medvram preset result in decent memory savings without huge performance hit
Doggetx 0.9 / 1.1 / 2.2 / 4.3 / 6.4 1.0 6.3 using lowvram preset is extremely slow due to constant swapping
xFormers 2.8 / 2.8 / 2.8 / 3.1 / 4.1 6.5 7.5 fastest and low memory
xFormers 2.9 / 2.9 / 2.9 / 3.6 / 4.1 6.4 7.6 with cuda_alloc_conf and opt-channelslast

Notes:

  • Performance at batch-size 1 is around ~70% of peak performance
  • Peak performance is typically around batch size 8
    After that it grows by few percent if you have extra VRAM before it starts to drop due to GC kicking in
  • Performance with lowvram preset is very low below batch size 8 and by then memory savings are not that big

Other possible optimizations:

  • adding set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512 in webui-user.bat
    No performance impact and increases initial memory footprint a bit but reduces memory fragmentation in long runs
  • opt-channelslast
    Hit-and-miss: seems like additional slight performance increase with higher batch sizes and slower with small sizes, but differences are within margin-of-error