3jlVSSKJr SSKJr SSKJrJr SSKrSSKJ r SSK J r J r SSK Jr SS KJrJrJr S S KJr \R*"\5rS r\"S S\55r"SS\5rSS/rg)) annotations) dataclass)AnyCallableN)tqdm)MultiPipelineCallbacksPipelineCallback)BlockRefinementScheduler) BaseOutputloggingreplace_example_docstring)DiffusionPipelinea  Examples: ```python >>> import torch >>> from transformers import AutoModelForCausalLM, AutoTokenizer >>> from diffusers import BlockRefinementScheduler, LLaDA2Pipeline >>> model_id = "inclusionAI/LLaDA2.1-mini" >>> model = AutoModelForCausalLM.from_pretrained( ... model_id, trust_remote_code=True, dtype=torch.bfloat16, device_map="auto" ... ) >>> tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) >>> scheduler = BlockRefinementScheduler() >>> pipe = LLaDA2Pipeline(model=model, scheduler=scheduler, tokenizer=tokenizer) >>> output = pipe(prompt="What is the meaning of life?", gen_length=256) >>> print(output.texts[0]) ``` c.\rSrSr%S\S'SrS\S'Srg)LLaDA2PipelineOutput5torch.LongTensor sequencesNlist[str] | Nonetexts)__name__ __module__ __qualname____firstlineno____annotations__r__static_attributes__rd/home/wildlama/miniconda3/lib/python3.13/site-packages/diffusers/pipelines/llada2/pipeline_llada2.pyrr5s"E "rrc^\rSrSr%SrS\S'S\S'S\S'/SQrSSU4S jjjr\S 5r SS jr SS jr \ R"5\"\5SSS jj55rSrU=r$)LLaDA2Pipeline;a Pipeline for LLaDA2-style discrete diffusion text generation via block-wise iterative refinement. This pipeline maintains a template sequence filled with a `mask_token_id` and refines it in blocks. In each refinement step, it samples candidate tokens for the active block and commits a subset based on confidence. The model is expected to accept an attention mask and `position_ids`, and to return logits of shape `[batch, seq, vocab_size]`. rmodelr scheduler tokenizer)block_xx0x0_ptransfer_index confidence active_blockc>[TU]5 URXUS9 URb[ URSS5OSUlURb[ URSS5UlgSUlg)N)r$r%r& eos_token_id mask_token_id)super__init__register_modulesr&getattrr.r/)selfr$r%r& __class__s r r1LLaDA2Pipeline.__init__Lsj  E)TMQ^^MgGDNNNDImqOS~~OiWT^^_dKosrcUR$N)_num_timesteps)r4s r num_timestepsLLaDA2Pipeline.num_timestepsWs"""rc"UbURS:XaURS5nURS:wa"[S[UR5S35eUR [ R:wa[SUR S35eU$URc [S5eUbUb [S 5eUcUc [S 5eU=(d 0nUb'URR"U4US S S S .UD6nUS$U(ag[URSS5(aK[U[5(a [S5eURR"SUS./4US S S S .UD6nUS$URUS [U[5S9nUS$)z?Convert prompt/messages/input_ids to a [batch, seq] LongTensor.Nrrz"`input_ids` must be 2D, got shape ./`input_ids` must be int64 token IDs, got dtype=7Tokenizer is required when `input_ids` is not provided.0Provide either `prompt` or `messages`, not both.4Provide one of `prompt`, `messages`, or `input_ids`.Tpt)add_generation_prompttokenizereturn_tensors return_dict input_ids chat_templatez8`prompt` must be a string when `use_chat_template=True`.user)rolecontent)rFpadding) ndim unsqueeze ValueErrortupleshapedtypetorchlongr&apply_chat_templater3 isinstancelist)r4promptmessagesrHuse_chat_templaterDchat_template_kwargsencodeds r _prepare_input_ids!LLaDA2Pipeline._prepare_input_ids]s  ~~"%//2 ~~" #EeIOOF\E]]^!_``%**, #RS\SbSbRccd!eff  >> !VW W  F$6OP P  ST T39r  nn88&;#  ' G;' ' $!O!O&$'' ![\\nn88 V45&;#  ' G;' '..jQWY]F^._{##rc ^UcUcUc [S5eUbUb [S5eUbiURS;a"[S[UR5S35eUR[ R :wa[SURS35eUbUcTRc [S5eUbUcTRc [S5eUS::a[S US35eUS::a[S US35eUS::a[S US35eUS::a[S US35eS Us=::aS::dO US:d[SUS35eU S;a[SU <S35eU S;a[SU <S35eU b'[U [[45(a U Rn U bX[U4SjU 55(d=[STRSU V s/sHoTR;dMU PM sn 35eggs sn f)NrBrA)r=rz(`input_ids` must be 1D or 2D, got shape r>r?r@rz`gen_length` must be > 0, got z `block_length` must be > 0, got z'`num_inference_steps` must be > 0, got z `minimal_topk` must be > 0, got g?zC`threshold` must be in [0, 1] (or > 1 to force top-k commits), got >autogreedy multinomialzF`sampling_method` must be one of {'auto','greedy','multinomial'}, got >seqtextz+`output_type` must be 'seq' or 'text', got c3@># UHoTR;v M g7fr8)_callback_tensor_inputs).0kr4s r .LLaDA2Pipeline.check_inputs..sF 7Y!-- -7Ysz2`callback_on_step_end_tensor_inputs` has to be in z , but found )rPrNrQrRrSrTrUr&rWr r tensor_inputsallrh)r4rYrZrH gen_length block_lengthnum_inference_steps minimal_topk thresholdsampling_method output_typecallback_on_step_end"callback_on_step_end_tensor_inputsrjs` r check_inputsLLaDA2Pipeline.check_inputssk >h.93DST T  ("6OP P  ~~V+ #KER[RaRaLbKccd!eff%**, #RS\SbSbRccd!eff  )"38NVW W  I$5$..:PVW W ?=j\KL L 1 ? ~QOP P ! #FGZF[[\]^ ^ 1 ? ~QOP Py'C')c/bclbmmnop p "C CZ[jZmmno  o -J;/YZ[\ \ + #35K"L1 1 2F1S1S . - 9#F 7YF C C DTEaEaDbbnAkA!dNjNjEjAAkln C 9 ls G9(G9c Ub'[U[[45(a URnUcS/nUR UUUUUUUU U UUUS9 UR UUUUUSS9nUR nURS:XaURS5nURUS9nURunnUc URnUc URnUc [S5e[XU-5nURR!UUS9 UU-U-S- U-nUU-n["R$"UU4U["R&S 9n["R("UU["R&S 9RS5R+US 5n ["R,"UU4UU["R&S 9n!US:a UU!SS2SU24'UU-n"U[/UU"- S5-Ul["R2"U4U["R4S 9n#USL=(a US :n$Sn%[7US 05R95n&SU&S 'SU&S'[;[=U"U540U&D6GHZn'U'S-U-n(U!SS2SU(24n)USS2SU(24n*U SS2SU(24n+U'U-n,["R2"UU["R4S 9n-U,U:a[UU,- U5n.SU-SU.&Sn/Sn0Sn1UR?SSSU'S3S9 URAUS9n2U1(Ga|U)SS2U*S24n3U3U:HRC5n4U4(dU/S- n/UREU)U*U+S9RFn5U5SS2U*S2SS24n6URRIU6U0U3UU U U U U UUU-USS9n7U7RJn8U7RLn9U8U9-n:U:RC5(aU7RNU)SS2U*S24'U(a,Ub)URRQU)U7RRU:U#UUUS9n#Ub80n;UHn<[U5U<U;U<'M U"UU%U0U;5n=U=RWSU)5n)U%S- n%U4(aU0S- n0U2RYS5 URR[U0U4U$U9U/UU#S9n1U1(aGM|U2R]5 U)U!SS2SU(24'U(dGMCU#R_5(dGM[ O U!SS2SUU-24n>U>SS2US24n?UbUUS:XaOU?SU:HRaSS9Sn@[cU@5S:a(U?SS2S[eW@SRg55S-24n?SnAUS:Xa'URhbURhRkU?SS9nAU(dU?RUS9WA4$[mU?RUS9WAS9$)a Generate text with block-wise refinement. Args: prompt (`str` or `List[str]`, *optional*): Prompt text. When `use_chat_template` is `True` (default) and a tokenizer with a chat template is available, the prompt is wrapped in a chat message before tokenization. messages (`List[Dict[str, str]]`, *optional*): Chat messages to encode (e.g. `[{"role": "user", "content": "Hello"}]`). Takes precedence over `prompt` when provided. Requires a tokenizer with `apply_chat_template`. input_ids (`torch.LongTensor`, *optional*): Pre-tokenized input IDs. Takes precedence over `prompt` and `messages`. use_chat_template (`bool`, defaults to `True`): Whether to wrap the prompt in a chat template. add_generation_prompt (`bool`, defaults to `True`): Whether to add the generation prompt when using chat templates. gen_length (`int`): Number of tokens to generate. block_length (`int`): Block size for refinement. num_inference_steps (`int`): Number of refinement steps per block. temperature (`float`): Sampling temperature. top_p (`float`, *optional*): Nucleus sampling cutoff. top_k (`int`, *optional*): Top-k sampling cutoff. sampling_method (`str`): Sampling method (`auto`, `greedy`, `multinomial`). threshold (`float`): Confidence threshold for committing tokens. editing_threshold (`float`, *optional*): Confidence threshold for editing already-committed (non-mask) tokens. When positive, after all mask tokens in a block are resolved, the pipeline continues refining: if the model predicts a different token with confidence above this threshold, the existing token is replaced. Set to `None`, `0.0`, or a negative value to disable editing. Defaults to `0.5`. max_post_steps (`int`): Maximum number of additional refinement iterations after all mask tokens in a block are resolved. Only used when `editing_threshold` is enabled. Defaults to `16`. minimal_topk (`int`): Minimum number of tokens to commit per step. eos_early_stop (`bool`): Whether to stop after committing EOS in a block. eos_token_id (`int`, *optional*): EOS token ID to use for early stopping. mask_token_id (`int`, *optional*): Mask token ID to use for the template. generator (`torch.Generator`, *optional*): RNG for sampling. output_type (`str`, defaults to `"text"`): Output format. `"text"` decodes sequences into strings (requires a tokenizer). `"seq"` returns raw token ID sequences only. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`LLaDA2PipelineOutput`] instead of a tuple. callback_on_step_end (`Callable` or `PipelineCallback`, *optional*): Callback executed after each refinement step with signature `callback_on_step_end(self, step: int, timestep: int, callback_kwargs: Dict)`. callback_on_step_end_tensor_inputs (`List[str]`, *optional*): Tensor keys to pass to the callback. Allowed keys: `block_x`, `x0`, `x0_p`, `transfer_index`, `confidence`, `active_block`. Examples: Nr') rYrZrHrorprqrrrsrtrurvrw)rYrZrHr[rDr\r=r)devicezA`mask_token_id` must be provided (or available on the tokenizer).)r{rSra_progress_bar_configpositionBlocksdescTFzBlock z Inference Steps)r~leaver)total)attention_mask position_ids) model_outputtimestepsampler/ temperaturetop_ptop_krtrsediting_thresholdrr prompt_mask generatorrG)cur_xsampled_tokensfinal_transferfinishedr.r/ prompt_length)step_idxmasks_remainingediting_enabledediting_transfer_index post_stepsmax_post_stepsr)as_tuplerf)skip_special_tokens)rr)7rWr r rmrxr^_execution_devicerNrOtorRr.r/rPminr% set_timestepsrTonesrUarangeexpandfullmaxr9zerosboolr3copyrrangeset_progress_bar_config progress_baranyr$logitsstepr*r prev_samplecheck_eos_finishedrlocalspopupdatecheck_block_should_continueclosernnonzerolenintitemr& batch_decoder)Br4rYrZrHr[rDrorprqrrrrtrsrrrreos_early_stopr.r/rrurGrvrw prompt_idsr{ batch_sizer num_blocks total_length attn_maskrxprefill_blocksrr global_stepblock_progress_bar_config num_blockcurrent_window_endr'block_attn_maskblock_position_idsblock_start_posprompt_mask_in_blockprompt_end_in_blockrrshould_continuer block_tokensrr block_logitsscheduler_outputr*rrcallback_kwargsrjcallback_outputs generatedr eos_positionsrsB r __call__LLaDA2Pipeline.__call__sB + #35K"L1 1 2F1S1S . - 52; . !% 3%+#!5/Q  ,,/"7!% - '' ??a #--a0J]]&]1 $.$4$4! M  ,,L   ..M  `a a!"5\7QR $$%8$H$j0<?!C T !L0 JJ L9&PUPZPZ[ ||LuzzR\\]^_ffgqsuv  JJ L1=W\WaWa b 1 #-Aa- &,61C ^8SUV4WW;; }V5::N+47S,>+>(>?O!-a1D2D1D.D!E (,6O#(;;|FRWR\R\#] .&)-/*I<&X#=A$%9&9:JH"O  ( (!5PY{ZjGk ( l,,3F,GL!&q<-.'89 #/=#@"E"E"G&!OJGOZlmtt%a,&:; #'>>#6#6!-%'"/ +$3'&7!- 4' $$7$ ""2!@!@)9)P)P&!/2H!H!%%''1A1M1MGA }~-.!l&>#~~@@%'7'F'F'5!)%1&3&3 A H(3&(O?-3Xa[*@';D+xYh'i$.229gFGq "MH ''*"&.."L"L%$3$3+A)#1%#M#s"/F    (/Aa$$$$ %~(,,..s^xa5=:5556 a/0  # a&q\\9BBDBQRSTM=!A%%a)K3}Q/?/D/D/F+G!+K)K&KL  & T^^%?NN// t/TE<rsy#! A2CC.   H %* #:# # mX&mX` 3 4r