--fast fp16_accumulation only helps with models that actually run in FP16, and most VAEs use BF16 or FP32, so this flag would have no effect on them. MiniMax-H3 VAE was a good target for this as it's slow and always used specifically FP16, and if any other VAE ever uses FP16 we can extend the optimization to that.
For BF16 VAEs there are other similar things we can do, image model decodes are usually so fast that it's not been a priority though.
Good reminder that encode/decode is part of wall-clock. It's why the MI355X figures (5.39s for a 15s clip, encoding excluded) shouldn't be lined up against end-to-end runs.
Thanks KJ. Shall we start ComfyUI with --fast fp16_accumulation for the faster encoder and decoder for Qwen Image 2.1, Krea2, Minimax Music H3, etc?
--fast fp16_accumulation only helps with models that actually run in FP16, and most VAEs use BF16 or FP32, so this flag would have no effect on them. MiniMax-H3 VAE was a good target for this as it's slow and always used specifically FP16, and if any other VAE ever uses FP16 we can extend the optimization to that.
For BF16 VAEs there are other similar things we can do, image model decodes are usually so fast that it's not been a priority though.
Awesome, thanks a lot Comfy team and a very special thanks to Kijai.
Good reminder that encode/decode is part of wall-clock. It's why the MI355X figures (5.39s for a 15s clip, encoding excluded) shouldn't be lined up against end-to-end runs.