add aoti for speed up

#5
by linoyts HF Staff - opened
LTX Community org

Loads a precompiled AOTI graph of the transformer blocks from linoyts/LTX-2.3-Native-Transformer-GroupA-sm120-cu130-r20 and patches the block loop to use it. bf16 (no quantization) — quality matches eager. ~2.5x warm on beard-removal. Also removes the runtime xformers install (pulls torch 2.8 -> breaks ZeroGPU torch 2.11) and pins requirements. NOTE: make linoyts/LTX-2.3-Native-Transformer-GroupA-sm120-cu130-r20 public (or add HF_TOKEN secret) before merging.

linoyts changed pull request title from AOTI: faster transformer (native bf16, ZeroGPU) to add aoti for speed up
LTX Community org

beard-removal merged demo builds & runs OK; merging AOTI per linoyts approval.

linoyts changed pull request status to merged

Sign up or log in to comment