9jFdZddlZddlZddlmZmZGddeZdS)z LCM + LTX scheduler combining Latent Consistency Model with RectifiedFlow (LTX). Optimized for Lightning LoRA compatibility and ultra-fast inference. N)SchedulerMixinSchedulerOutputcveZdZdZddededefdZddedefd Zd ej d ej d ej de fdZ dZ d S) LCMSchedulera LCM + LTX scheduler combining Latent Consistency Model with RectifiedFlow. - LCM: Enables 2-8 step inference with consistency models - LTX: Uses RectifiedFlow for better flow matching dynamics Optimized for Lightning LoRAs and ultra-fast, high-quality generation. ?num_train_timestepsnum_inference_stepsshiftc>||_||_||_d|_dS)N)r r r _step_index)selfr r r s BC:\Users\johnn\OneDrive\Documents\AI\shared\utils\lcm_scheduler.py__init__zLCMScheduler.__init__s'#6 #6  Nc t|d|_||j}tjdd|jdztj|}d}d}|||z d|z zz}||zd|dz |zzz }||_|jdd|jz|_|>|j ||_|j ||_d|_ dS) z@Set timesteps for LCM+LTX inference using RectifiedFlow approachNr)dtypedevicer gJh?) minr r torchlinspacefloat32sigmasr timestepstor) rr rr kwargst sigma_max sigma_minrs r set_timestepszLCMScheduler.set_timestepss#&':A#>#> =JE N1a!9A!=U][a b b b ! i)3A>>1 V';#;< SbS)D,DD  !^..v66DN+..00DKr model_outputtimestepsamplereturnc |j|||j|j}|jdzt|jkr|j|jdz}nt j|}||z }t|jt|jkr?|d}t|jt|jk?|||zz}|xjdz c_t|S)z Perform LCM + LTX step combining consistency model with rectified flow. - LCM: Direct consistency model prediction for fast inference - LTX: RectifiedFlow dynamics for optimal probability flow path Nrr) prev_sample) r_init_step_indexrlenr zeros_likeshape unsqueezer) rr%r&r'r sigma sigma_next sigma_diffr*s rstepzLCMScheduler.step8s    #  ! !( + + + D,-  a #dk"2"2 2 2T%5%9:JJ)%00J!5( *"##c&,&7&777#--b11J*"##c&,&7&777 |j88  A;7777rct|tjr||jj}|j|k}t|dkr!|d|_ dStj |j|z }tj ||_ dS)z/Initialize step index based on current timesteprN) isinstancerTensorrrrnonzeror,itemrabsargmin)rr&indicesdiffss rr+zLCMScheduler._init_step_indexYs h - - :{{4>#899H>X-6688 w<888u|8jy8888B : : : : :rr)r@rmath%diffusers.schedulers.scheduling_utilsrrrrCrrrFsy  QQQQQQQQX:X:X:X:X:>X:X:X:X:X:r