bi|xddlZddlmZmZmZmZddlZddlZddl m Z m Z ddl m Z ddlmZmZGdd ee Zy) N)ListOptionalTupleUnion) ConfigMixinregister_to_config) randn_tensor)SchedulerMixinSchedulerOutputc#eZdZdZgZdZe d9dedededede ded ed e d e d ed ededede de de ef dZ e dZe dZe dZd:de fdZdZdZdZdej,deeej,fdej,fdZd;d!e d"eeej2ffd#Zd;dej,fd$Zd;dej,fd%Zdej,dej,fd&Zd'Zd(Z dEDMDPMSolverMultistepSchedulera% Implements DPMSolverMultistepScheduler in EDM formulation as presented in Karras et al. 2022 [1]. `EDMDPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs. [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models." https://huggingface.co/papers/2206.00364 This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic methods the library implements for all schedulers such as loading and saving. Args: sigma_min (`float`, *optional*, defaults to 0.002): Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable range is [0, 10]. sigma_max (`float`, *optional*, defaults to 80.0): Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable range is [0.2, 80.0]. sigma_data (`float`, *optional*, defaults to 0.5): The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1]. sigma_schedule (`str`, *optional*, defaults to `karras`): Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl. num_train_timesteps (`int`, defaults to 1000): The number of diffusion steps to train the model. solver_order (`int`, defaults to 2): The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided sampling, and `solver_order=3` for unconditional sampling. prediction_type (`str`, defaults to `epsilon`, *optional*): Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) paper). thresholding (`bool`, defaults to `False`): Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such as Stable Diffusion. dynamic_thresholding_ratio (`float`, defaults to 0.995): The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. sample_max_value (`float`, defaults to 1.0): The threshold value for dynamic thresholding. Valid only when `thresholding=True` and `algorithm_type="dpmsolver++"`. algorithm_type (`str`, defaults to `dpmsolver++`): Algorithm type for the solver; can be `dpmsolver++` or `sde-dpmsolver++`. The `dpmsolver++` type implements the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. solver_type (`str`, defaults to `midpoint`): Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. lower_order_final (`bool`, defaults to `True`): Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. euler_at_final (`bool`, defaults to `False`): Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference steps, but sometimes may result in blurring. final_sigmas_type (`str`, defaults to `"zero"`): The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. r sigma_min sigma_max sigma_datasigma_schedulenum_train_timestepsprediction_typerho solver_order thresholdingdynamic_thresholding_ratiosample_max_valuealgorithm_type solver_typelower_order_finaleuler_at_finalfinal_sigmas_typec| dvr2| dk(r|jdnt| d|j| dvr1| dvr|jd nt| d|j| dvr|d k(rtd |d | d t j dd|}|dk(r|j |}n|dk(r|j|}|j|_ t j|t jd|jg|_ d|_dg|z|_d|_d|_d|_|jj'd|_ y)N) dpmsolver++sde-dpmsolver++deisr!)rz is not implemented for )midpointheun)logrhobh1bh2r$)rzeroz`final_sigmas_type` z' is not supported for `algorithm_type` z$. Please choose `sigma_min` instead.rr karras exponential)devicecpu)r NotImplementedError __class__ ValueErrortorchlinspace_compute_karras_sigmas_compute_exponential_sigmasprecondition_noise timestepscatzerosr,sigmasnum_inference_steps model_outputslower_order_nums _step_index _begin_indexto)selfrrrrrrrrrrrrrrrrrampr9s r/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py__init__z'EDMDPMSolverMultistepScheduler.__init__[s* !C C'''}'E)^,<rHs rB begin_indexz*EDMDPMSolverMultistepScheduler.begin_indexs    rDrNc||_y)z Sets the begin index for the scheduler. This function should be run from pipeline before the inference. Args: begin_index (`int`): The begin index for the scheduler. NrM)r@rNs rBset_begin_indexz.EDMDPMSolverMultistepScheduler.set_begin_indexs (rDc2|j|}||z}|SN)_get_conditioning_c_in)r@samplesigmac_in scaled_samples rBprecondition_inputsz2EDMDPMSolverMultistepScheduler.precondition_inputss!**51 rDct|tjstj|g}dtj|z}|S)Ng?) isinstancer1Tensortensorlog)r@rUc_noises rBr5z1EDMDPMSolverMultistepScheduler.precondition_noises8%.LL%)E5))rDcj|jj}|dz|dz|dzzz }|jjdk(r||z|dz|dzzdzz }nR|jjdk(r| |z|dz|dzzdzz }n#td|jjd||z||zz}|S)NrepsilonrF v_predictionzPrediction type z is not supported.)rGrrr0)r@rT model_outputrUrc_skipc_outdenoiseds rBprecondition_outputsz3EDMDPMSolverMultistepScheduler.precondition_outputss[[++ Q%(Z]":; ;; & &) 3J&%(Z]*Bs)JJE [[ ( (N :FZ'5!8j!m+C*KKE/ 0K0K/LL^_` `F?U\%99rDrTtimestepreturnc|j|j||j|j}|j||}d|_|S)a Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. Args: sample (`torch.Tensor`): The input sample. timestep (`int`, *optional*): The current timestep in the diffusion chain. Returns: `torch.Tensor`: A scaled input sample. T)rK_init_step_indexr9rXis_scale_input_called)r@rTrgrUs rBscale_model_inputz0EDMDPMSolverMultistepScheduler.scale_model_inputsM ?? "  ! !( + DOO,))&%8%)" rDNr:r,cr||_tjdd|j}|jjdk(r|j |}n*|jjdk(r|j |}jtj|}|j||_ |jjdk(r|jj}n>|jjdk(rd}n"td|jjtj|tj|gtj|g|_d g|jj"z|_d|_d |_d |_|j jd |_y ) a Sets the discrete timesteps used for the diffusion chain (to be run before inference). Args: num_inference_steps (`int`): The number of diffusion steps used when generating samples with a pre-trained model. device (`str` or `torch.device`, *optional*): The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. rr r*r+)dtyper,rr)zC`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got Nr-)r:r1r2rGrr3r4r?float32r5r6rrr0r7r\r9rr;r<r=r>)r@r:r,rAr9 sigma_lasts rB set_timestepsz,EDMDPMSolverMultistepScheduler.set_timestepssi$7 ~~aD$<$<= ;; % % 1006F [[ ' '= 855d;Fv>008 ;; ( (K 7..J [[ * *f 4JUVZVaVaVsVsUtu iizl%--`f)g hi   KK $ $%!"  kknnU+ rDc|xs|jj}|xs|jj}|jj}|d|z z}|d|z z}||||z zz|z}|S)z6Constructs the noise schedule of Karras et al. (2022).r )rGrrr)r@rArrr min_inv_rho max_inv_rhor9s rBr3z5EDMDPMSolverMultistepScheduler._compute_karras_sigmassq6!6!6 6!6!6 kkooAG, AG,  k(A BBsJ rDc4|xs|jj}|xs|jj}tjt j |t j |t|jjd}|S)zImplementation closely follows k-diffusion. https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 r) rGrrr1r2mathr]lenexpflip)r@rArrr9s rBr4z:EDMDPMSolverMultistepScheduler._compute_exponential_sigmas!sn 6!6!6 6!6!6  3TXXi5H#d)TXXZ__`ab rDcb|j}|j^}}}|tjtjfvr|j }|j ||tj|z}|j}tj||jjd}tj|d|jj}|jd}tj|| ||z }|j ||g|}|j!|}|S)a{ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing pixels from saturation at each step. We find that dynamic thresholding results in significantly better photorealism as well as better image-text alignment, especially when using very large guidance weights." https://huggingface.co/papers/2205.11487 r )dim)minmax)rnshaper1rofloat64floatreshapenpprodabsquantilerGrclampr unsqueezer?)r@rTrn batch_sizechannelsremaining_dims abs_sampless rB_threshold_samplez0EDMDPMSolverMultistepScheduler._threshold_sample,s 06 - H~  6 6\\^F Hrww~7N,NOZZ\ NN:t{{'M'MST U KK 1$++66  KKNVaR+a/ HF~F5! rDctjtj|d}||ddtjfz }tj|dk\dj dj |jddz }|dz}||}||}||z ||z z } tj | dd} d| z |z| |zz} | j|j} | S)Ng|=r)axisr)r}r ) rr]maximumnewaxiscumsumargmaxclipr~r) r@rU log_sigmas log_sigmadistslow_idxhigh_idxlowhighwts rB _sigma_to_tz*EDMDPMSolverMultistepScheduler._sigma_to_tNsFF2::eU34 Jq"**}55))UaZq188a8@EE*JZJZ[\J]`aJaEbQ;!(#9_t , GGAq! Ug H , IIekk "rDc8tjd}|}||fS)Nr )r1r\)r@rUalpha_tsigma_ts rB_sigma_to_alpha_sigma_tz6EDMDPMSolverMultistepScheduler._sigma_to_alpha_sigma_tes,,q/rDrbc|j|j}|j|||}|jjr|j |}|S)a0 Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an integral of the data prediction model. The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise prediction and data prediction models. Args: model_output (`torch.Tensor`): The direct output from the learned diffusion model. sample (`torch.Tensor`): A current instance of a sample created by the diffusion process. Returns: `torch.Tensor`: The converted model output. )r9rKrfrGrr)r@rbrTrUx0_preds rBconvert_model_outputz3EDMDPMSolverMultistepScheduler.convert_model_outputksL6 DOO,++FL%H ;; # #,,W5GrDnoisec "|j|jdz|j|j}}|j|\}}|j|\}}tj|tj|z }tj|tj|z } || z } |j j dk(r*||z |z|tj| dz z|zz } | S|j j dk(rz|J||z tj| z|z|dtjd| zz z|zz|tjdtjd| zz z|zz}  S)a One step for the first-order DPMSolver (equivalent to DDIM). Args: model_output (`torch.Tensor`): The direct output from the learned diffusion model. sample (`torch.Tensor`): A current instance of a sample created by the diffusion process. Returns: `torch.Tensor`: The sample tensor at the previous timestep. r r!?r") r9rKrr1r]rGrrxsqrt) r@rbrTrrsigma_sralpha_slambda_tlambda_shx_ts rBdpm_solver_first_order_updatez'>??%GH   rDmodel_output_listc Z|j|jdz|j|j|j|jdz }}}|j|\}}|j|\}}|j|\} }tj|tj|z } tj|tj|z } tj| tj|z } |d|d}} | | z | | z }}||z }| d|z | |z z}}|j j dk(r|j jdk(rM||z |z|tj| dz z|zz d|tj| dz zz|zz }|S|j jdk(rN||z |z|tj| dz z|zz |tj| dz |z dzz|zz}S|j j d k(rt|J|j jdk(r||z tj| z|z|dtjd |zz z|zzd|dtjd |zz zz|zz|tjdtjd|zz z|zz}|S|j jdk(r||z tj| z|z|dtjd |zz z|zz|dtjd |zz d |zz dzz|zz|tjdtjd|zz z|zz}S) a One step for the second-order multistep DPMSolver. Args: model_output_list (`List[torch.Tensor]`): The direct outputs from learned diffusion model at current and latter timesteps. sample (`torch.Tensor`): A current instance of a sample created by the diffusion process. Returns: `torch.Tensor`: The sample tensor at the previous timestep. r rrr!r$rFr%r"r) r9rKrr1r]rGrrrxr)r@rrTrrsigma_s0sigma_s1ralpha_s0alpha_s1r lambda_s0 lambda_s1m0m1rh_0r0D0D1rs rB(multistep_dpm_solver_second_order_updatezGEDMDPMSolverMultistepScheduler.multistep_dpm_solver_second_order_updates( KK!+ , KK ( KK!+ ,$  77@!99(C(!99(C(99W% '(::IIh'%))H*== IIh'%))H*== "2&(9"(=BI%y9'<3 1WcBh27+B ;; % % 6{{&&*4x'61%))QB-#"56"<=W 1" (;<=BC8 /((F2x'61%))QB-#"56"<=599aR=3#6!";c"ABbHI, #[[ ' '+< <$ $${{&&*4x'%))QB-76A!eiiq&9"9:b@AWEIIdQh,?(?@ABFG 3261B+B CCeKL ((F2x'%))QB-76A!eiiq&9"9:b@A34!8)<#<"JS"PQUWWX 3261B+B CCeKL rDc|j|jdz|j|j|j|jdz |j|jdz f\}}}}|j|\}}|j|\}}|j|\} }|j|\} }tj|tj|z } tj|tj|z } tj| tj|z } tj| tj|z }|d|d|d}}}| | z | | z | |z }}}||z ||z }}|}d|z ||z zd|z ||z z}}||||zz ||z zz}d||zz ||z z}|j j dk(rz||z |z|tj| dz z|zz |tj| dz |z dzz|zz|tj| dz |z|dzz dz z|zz }S) a One step for the third-order multistep DPMSolver. Args: model_output_list (`List[torch.Tensor]`): The direct outputs from learned diffusion model at current and latter timesteps. sample (`torch.Tensor`): A current instance of a sample created by diffusion process. Returns: `torch.Tensor`: The sample tensor at the previous timestep. r rrrrr!rF)r9rKrr1r]rGrrx)r@rrTrrrsigma_s2rrralpha_s2rrr lambda_s2rrm2rrh_1rr1rD1_0D1_1rD2rs rB'multistep_dpm_solver_third_order_updatezFEDMDPMSolverMultistepScheduler.multistep_dpm_solver_third_order_updates& KK!+ , KK ( KK!+ , KK!+ , 1 -8X 77@!99(C(!99(C(!99(C(99W% '(::IIh'%))H*== IIh'%))H*== IIh'%))H*== &r*,=b,ACTUWCXB*I ,A9yCX3q#'B Bh27+cBh27-Cd R27^t 4 4R"Wo$+ . ;; % % 68#v-eiimc12b89uyy!}s2a7#=>"DEuyy!}s2Q6!Q$>DEKL  rDc| |j}||k(j}t|dk(rt|jdz }|St|dkDr|dj}|S|dj}|S)Nrr )r6nonzerorwitem)r@rgschedule_timestepsindex_candidatesrKs rBindex_for_timestepz1EDMDPMSolverMultistepScheduler.index_for_timestep2s  %!% .(:CCE  A %T^^,q0J ! "Q &)!,113J*!,113JrDc|jVt|tjr%|j |j j }|j||_y|j|_y)zF Initialize the step_index counter for the scheduler. N) rNrZr1r[r?r6r,rr=r>)r@rgs rBrjz/EDMDPMSolverMultistepScheduler._init_step_indexFsW    #(ELL1#;;t~~'<'<=#66x@D #00D rD return_dictc|j td|j|j||jt |j dz k(xrc|j jxsK|j jxrt |j dkxs|j jdk(}|jt |j dz k(xr0|j jxrt |j dk}|j||}t|j jdz D]!}|j|dz|j|<#||jd<|j jd k(r.t|j ||j"|j$ } nd} |j jdk(s|j&dks|r|j)||| } nf|j jdk(s|j&dks|r|j+|j|| } n|j-|j|} |j&|j jkr|xj&dz c_|xj.dz c_|s| fSt1| S) a Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with the multistep DPMSolver. Args: model_output (`torch.Tensor`): The direct output from learned diffusion model. timestep (`int`): The current discrete timestep in the diffusion chain. sample (`torch.Tensor`): A current instance of a sample created by the diffusion process. generator (`torch.Generator`, *optional*): A random number generator. return_dict (`bool`): Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. Returns: [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a tuple is returned where the first element is the sample tensor. NzaNumber of inference steps is 'None', you need to run 'set_timesteps' after creating the schedulerr r)r)rTrr") generatorr,rn)rTr) prev_sample)r:r0rKrjrwr6rGrrrrrangerr;rr r~r,rnr<rrrr=r ) r@rbrgrTrrrlower_order_secondirrs rBstepz#EDMDPMSolverMultistepScheduler.stepRs<  # # +s  ?? "  ! !( +"__DNN0Ca0GG KK & & 7 --J#dnn2E2J 7{{,,6  __DNN 3a 7 7 wT[[=Z=Z w_bcgcqcq_ruw_w 00f0M t{{//!34 >A$($6$6q1u$=D  q ! >!-2 ;; % %): : ""i @S@S[g[m[mEE ;; # #q (D,A,AA,EIZ<<\RX`e<fK [[ % % *d.C.Ca.GK]GGHZHZciqvGwKFFtGYGYbhFiK  4;;#;#; ;  ! !Q & ! A> !;77rDoriginal_samplesr6c|jj|j|j}|jjdk(rvt j |ra|jj|jt j}|j|jt j}n@|jj|j}|j|j}|j |Dcgc]}|j||}}nG|j|jg|jdz}n|jg|jdz}||j}t|jt|jkr=|jd}t|jt|jkr=|||zz} | Scc}w)N)r,rnmps)rnrr)r9r?r,rntyper1is_floating_pointr6rorNrrKr~flattenrwr) r@rrr6r9rr step_indicesrU noisy_sampless rB add_noisez(EDMDPMSolverMultistepScheduler.add_noises'7'>'>FVF\F\]  " " ' '5 0U5L5LY5W!%!2!23C3J3JRWR_R_!2!` ! %5%<%rs/" // A,=l /^[l /rD