Clarification on 2.5× speedup baseline: base model or Turbo LoRA?
It delivers approximately 2.5× inference acceleration on an NVIDIA RTX 5090.
Could you clarify which baseline this 2.5× speedup is measured against — the base model, or the Turbo LoRA variant?
Thanks!
I think it's the lora that allows SLA, but no SLA support yet on Comfyui, you need to patch the model, AFAIK there is no custom nodes available to download to enable SLA, have to wait or make one yourself.
That still doesn't answer my question. Let me be very specific: is the 2.5× speedup measured against https://huggingface.co/lightx2v/Minimax-h3-Turbo, or against the base model?
A concrete example — suppose the base model takes 20s per inference and Minimax-h3-Turbo takes 10s. How long does this SLA Turbo LoRA take?
- 8s → 2.5× over the base model (making it just an alternative speedup method)
- 4s → 2.5× over Minimax-h3-Turbo (stacking on top of it)
Which one is it?
That still doesn't answer my question. Let me be very specific: is the 2.5× speedup measured against https://huggingface.co/lightx2v/Minimax-h3-Turbo, or against the base model?
A concrete example — suppose the base model takes 20s per inference and Minimax-h3-Turbo takes 10s. How long does this SLA Turbo LoRA take?
- 8s → 2.5× over the base model (making it just an alternative speedup method)
- 4s → 2.5× over Minimax-h3-Turbo (stacking on top of it)
Which one is it?
it's speeds up the S/IT (less seconds/IT), non-sla turbo loras do not speed that up, just lower required steps.
https://huggingface.co/lightx2v/Minimax-h3-Turbo-SLA/discussions/1#6a871e6c6c92fd2cdb074f10
His went from 44s/it to 25s/it, now this is from a comparison using Comfy kitchen attention for the 44s/IT, SLA with his custom made node was 25s/IT
Got it. Thanks
That still doesn't answer my question. Let me be very specific: is the 2.5× speedup measured against https://huggingface.co/lightx2v/Minimax-h3-Turbo, or against the base model?
A concrete example — suppose the base model takes 20s per inference and Minimax-h3-Turbo takes 10s. How long does this SLA Turbo LoRA take?
- 8s → 2.5× over the base model (making it just an alternative speedup method)
- 4s → 2.5× over Minimax-h3-Turbo (stacking on top of it)
Which one is it?
it's speeds up the S/IT (less seconds/IT), non-sla turbo loras do not speed that up, just lower required steps.
https://huggingface.co/lightx2v/Minimax-h3-Turbo-SLA/discussions/1#6a871e6c6c92fd2cdb074f10
His went from 44s/it to 25s/it, now this is from a comparison using Comfy kitchen attention for the 44s/IT, SLA with his custom made node was 25s/IT
Bruh...it requires les steps cause it's a 4 step turbo lora lol
That still doesn't answer my question. Let me be very specific: is the 2.5× speedup measured against https://huggingface.co/lightx2v/Minimax-h3-Turbo, or against the base model?
A concrete example — suppose the base model takes 20s per inference and Minimax-h3-Turbo takes 10s. How long does this SLA Turbo LoRA take?
- 8s → 2.5× over the base model (making it just an alternative speedup method)
- 4s → 2.5× over Minimax-h3-Turbo (stacking on top of it)
Which one is it?
it's speeds up the S/IT (less seconds/IT), non-sla turbo loras do not speed that up, just lower required steps.
https://huggingface.co/lightx2v/Minimax-h3-Turbo-SLA/discussions/1#6a871e6c6c92fd2cdb074f10
His went from 44s/it to 25s/it, now this is from a comparison using Comfy kitchen attention for the 44s/IT, SLA with his custom made node was 25s/IT
Bruh...it requires les steps cause it's a 4 step turbo lora lol
Turbo lora's generally do not lower the S/IT per step, just a lot better per step, so lower steps needed. This SLA lora not only reduces the amount of steps needed to get an acceptable quality generation, but it also speeds up inference, thanks to the attention mechanism/backend.
that is really crazy! look forward to the r2va! amazing work!