3j MSSKJr SSKJr SSKrSSKJrJr SSKJ r SSK J r \"S S \ 55r "S S \ \5r S S /rg) ) annotations) dataclassN) ConfigMixinregister_to_config) BaseOutput)SchedulerMixincL\rSrSr%SrS\S'S\S'S\S'S\S'S \S 'S rg ) BlockRefinementSchedulerOutputa Output class for block refinement scheduling. Args: prev_sample (`torch.LongTensor` of shape `(batch_size, block_length)`): Updated block tokens after the current refinement step. transfer_index (`torch.BoolTensor` of shape `(batch_size, block_length)`): Boolean mask indicating which tokens were committed (mask-filling). editing_transfer_index (`torch.BoolTensor` of shape `(batch_size, block_length)`): Boolean mask indicating which tokens were edited (non-mask replacement). sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`): Sampled token IDs from the model logits. sampled_probs (`torch.Tensor` of shape `(batch_size, block_length)`): Probabilities of the sampled tokens. torch.LongTensor prev_sampletorch.BoolTensortransfer_indexediting_transfer_indexsampled_tokens torch.Tensor sampled_probsN)__name__ __module__ __qualname____firstlineno____doc____annotations____static_attributes__rj/home/wildlama/miniconda3/lib/python3.13/site-packages/diffusers/schedulers/scheduling_block_refinement.pyr r s' "!$$,,$$rr c \rSrSrSrSr\SSSjj5rSSSjjrSSjr \ SSj5r \ SS j5r \ SS j5r S SSS SSSSSS S. SSjjr\ SSj5rSSjrSS.S SjjrSrg)!BlockRefinementScheduler3a( Scheduler for block-wise iterative refinement (commit-by-confidence). At each step, the scheduler samples candidate tokens from model logits and commits those with the highest confidence. The number of tokens to commit per step is determined by evenly distributing the block length across the number of refinement steps. Optionally supports editing: after all mask tokens are resolved, tokens can be replaced if the model predicts a different token with confidence above a positive `editing_threshold` (`None`, `0.0`, or negative disables editing). r NcX l[R"URS- SS[RS9UlSUlg)Nr dtype)num_inference_stepstorcharangelong timesteps_transfer_schedule)self block_lengthr' thresholdediting_threshold minimal_topks r__init__!BlockRefinementScheduler.__init__As:$7 d&>&>&BBRWR\R\];?rc8US::a[SUS35eXl[R"URS- SSU[RS9UlUR URRUR5RUbUOSS9Ul g) Nrz'`num_inference_steps` must be > 0, got .r r$devicer&cpu)r7) ValueErrorr'r(r)r*r+get_num_transfer_tokensconfigr.tor,)r-r'r7s r set_timesteps&BlockRefinementScheduler.set_timestepsNs ! #FGZF[[\]^ ^#6 d&>&>&BBSYafakakl"&">">t{{?W?WY]YqYq"r"u"u#/6U#v# rcUS::a#[R"S[RS9$X-nX-n[R"U4U[RS9nUSU===S- sss&U$)zREvenly distribute `block_length` token commits across `num_inference_steps` steps.r)rr%Nr )r(zerosr*full)r-r.r'base remainderouts rr:0BlockRefinementScheduler.get_num_transfer_tokensWs^ ! #;;t5::6 62 6 jj-/UZZH JY1 rcUbUS:aU$SUs=:aS::dO [SUS35e[R"USSS9up#[R"USS 9nUR SS 9nU[ U5:nUS SS24R 5US S S24'S US 'URU[R"UR5R5nURSX25nU$)z Nucleus (top-p) logit filtering.N?z`top_p` must be in (0, 1], got r5Tr$) descendingdimrJ.r r).r) r9r(sortsoftmaxcumsumfloatclone masked_fillfinfor&minscatter)logitstop_p sorted_logitssorted_indices sorted_probscumulative_probssorted_indices_to_removefiltereds r_top_p_filtering)BlockRefinementScheduler._top_p_filteringcs =ESLMe"s">ugQGH H(- 6dPR(S% }}]; '..2.6#3eEl#B ,DS#2#X,N,T,T,V ab)+, (%112JEKKXeXkXkLlLpLpq >>"nDrcUbUS::aU$XRS:aU$[R"XSS9up#USnURX:[R"UR 5R 5$)zTop-k logit filtering.rr$)krJ).r$N)shaper(topkrQrRr&rS)rUtop_kvalues_min_keeps r_top_k_filtering)BlockRefinementScheduler._top_k_filteringwsm =EQJM LL$ $MJJvB7 -(!!&"3U[[5N5R5RSSrcUS:a[SUS35eURSnURSU5nUS:XdU(d[R"UR 5SS9nUR SSS9n [R"USU 5n U R"URS S6U R"URS S64$Un US :waXq- n [RXS 9n [RXS 9n [R"U R 5SS9n[R"US US9n [R"USU 5n U R"URS S6U R"URS S64$)zESample tokens from logits with temperature scaling, top-k, and top-p.rz `temperature` must be >= 0, got r5r$rHrKT)rJkeepdimNrG)rc)rVr ) num_samples generator) r9rareshaper(rMrOargmaxgatherviewr!rgr] multinomial) rU temperaturercrVrluse_multinomial vocab_size flat_logitsprobstoken token_probscaledr\s r_sample_from_logits,BlockRefinementScheduler._sample_from_logitssh ?? }ANO O\\"% nnR4 # _MM+"3"3"52>E&&2t&>'='A'A'C'C x+C+CC:#>r)rlcURupxURn UR5n UR5n [R"U[R S9n [R"U[R S9n UR [R S9n[X8U5Hn[XU-5nUU- nUS::aM[R"US4XS9n[R"UU4XS9U:nUUSS2UU24-nU)USS2UU24-nUU SS2UU24'UU SS2UU24'M [R"U [R"X5U 5n [R"U [R"X5U 5n XX4$)uE Apply the forward (noising) process for semi-autoregressive block masking. For each block after the prompt, a random fraction of valid (non-padding) tokens are replaced with `mask_token_id`. Two complementary views are returned: `noisy` and `noisy_rev`, where the masked positions in one are the unmasked positions in the other. Args: original_samples (`torch.LongTensor` of shape `(batch_size, seq_len)`): Clean token IDs. attention_mask (`torch.LongTensor` of shape `(batch_size, seq_len)`): Padding mask (1 for valid, 0 for padding). prompt_length (`int`): Number of leading prompt tokens to keep unmasked. block_length (`int`): Block size for masking. mask_token_id (`int`): Token ID to use for masked positions. generator (`torch.Generator`, *optional*): RNG for reproducibility. Returns: `tuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor, torch.BoolTensor]`: `(noisy, noisy_rev, masked, masked_rev)` — the two complementary noisy sequences and their corresponding boolean masks. r%rr )r7rlN) rar7rPr(rrr<rrSrandrr)r-original_samplesattention_maskrr.rrlrseq_lenr7noisy noisy_revmasked masked_revvalid block_start block_endseg_lenp_masksegseg_revs r add_noise"BlockRefinementScheduler.add_noisesH/44 !(( &&($**, !!"2%**E%%&6ejjI !! !3 FKG<%?@I+-G!|ZZQTF**j'26WZ``CaY!6677CtuQ I(=%=>>G/2F1k)++ ,3:Jq+i// 0G FEOOE$I5QKK EOOI,UW`a 33r)r,r'r+) rgffffff?Nr ) r.rr'rr/rOr0 float | Noner1r)N)r'rr7zstr | torch.device | NonereturnNone)r.rr'rrr)rUrrVrrr)rUrrc int | Nonerr)rUrrrrOrcrrVrrltorch.Generator | Nonersrrz%tuple[torch.LongTensor, torch.Tensor])rrrzint | torch.TensorrrrrrrrOrVrrcrr}strr/rr0rr1rr~ztorch.BoolTensor | Nonerlrrrrz|BlockRefinementSchedulerOutput | tuple[torch.LongTensor, torch.BoolTensor, torch.BoolTensor, torch.LongTensor, torch.Tensor])rrrrrrrrrrrrrrrr)rrrrrrrrrrrrrrrr)rrrrrrr.rrrrlrrzMtuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor, torch.BoolTensor])rrrrrorderrr2r=r: staticmethodr]rgrzrrrrrrrrr!r!3s  E#%*. @ @! @ @ ( @  @ @ &TT!S!S!S !S  !S * !S!S /!S!ST!" %"&*.#'/3,0 !G "G %G ! G  G G G G G  G (G !G -G *G !G $ f%G R..(.).# .  .  .. ..`*** * !1 *  **#* *h-1=4*=4)=4  =4  =4=4*=4 W=4=4rr!) __future__r dataclassesrr(configuration_utilsrrutilsrscheduling_utilsr r r!__all__rrrrsR#! A,  Z   0V4~{V4r &'G Hr