i8;ddlmZddlZddlmZddlmZddlmZm Z ddl m Z m Z dd l mZmZd d lmZej"eZGd d eZGddeZGddeZGddeZy))ListN) FluxPipeline)logging)ModularPipelineBlocks PipelineState) InputParam OutputParam) calculate_dimension_from_latentsrepeat_tensor_to_batch_size)FluxModularPipelineceZdZdZedefdZedeefdZ edeefdZ dZ e jdededefd Zy ) FluxTextInputStepfluxreturnc y)NzText input processing step that standardizes text embeddings for the pipeline. This step: 1. Determines `batch_size` and `dtype` based on `prompt_embeds` 2. Ensures all text embeddings have consistent batch sizes (batch_size * num_images_per_prompt)selfs q/home/obispo/Crisostomo_bridge/mision_env/lib/python3.12/site-packages/diffusers/modular_pipelines/flux/inputs.py descriptionzFluxTextInputStep.description#s  p ctddtdddtjdtd dtjd gS) Nnum_images_per_promptr)default prompt_embedsTdenoiser_input_fieldszGPre-generated text embeddings. Can be generated from text_encoder step.)required kwargs_type type_hintrpooled_prompt_embedszNPre-generated pooled text embeddings. Can be generated from text_encoder step.)r!r"r)r torchTensorrs rinputszFluxTextInputStep.inputs,sN . : 3,,e   &3,,l    rc tdtdtdtjdtdtjdd td tjdd gS) N batch_sizezdNumber of prompts, the final batch size of model inputs should be batch_size * num_images_per_prompt)r"rdtypez@Data type of model tensor inputs (determined by `prompt_embeds`)rrz2text embeddings used to guide the image generation)r"r!rr#z9pooled text embeddings used to guide the image generation)r intr$r)r%rs rintermediate_outputsz&FluxTextInputStep.intermediate_outputs@sm C  ++^  ,,3P   &,,3W  #  rc|j{|jn|jjd|jjdk7r:td|jjd|jjdyyy)Nrzx`prompt_embeds` and `pooled_prompt_embeds` must have the same batch size when passed directly, but got: `prompt_embeds` z != `pooled_prompt_embeds` .)rr#shape ValueError)r components block_states r check_inputszFluxTextInputStep.check_inputs\s  $ $ 0[5U5U5a((..q1[5U5U5[5[\]5^^ --8-F-F-L-L,MN#88>>?qB_6b 0rr0statec|j|}|j|||jjd|_|jj |_|jj\}}}|jj d|jd|_|jj|j|jz|d|_|jj d|j}|j|j|jzd|_ |j||||fS)Nrr) get_block_stater2rr.r(r)repeatrviewr#set_block_state)rr0r3r1_seq_lenr#s r__call__zFluxTextInputStep.__call__es%**51  *k2!,!:!:!@!@!C '55;; #1177 7A$/$=$=$D$DQ HiHikl$m !$/$=$=$B$B  " "[%F%F FQS% ! +??FFq+JkJkl+?+D+D  " "[%F%F F, ( UK05  rN)__name__ __module__ __qualname__ model_namepropertystrrrr r&r+r2r$no_gradrr r<rrrrr sJ  S   Z(  & d3i  6U]]_!#6!}!Q^!!rrceZdZdZdggfdeedeeffd ZedefdZedee fdZ edee fd Z d e d edefd ZxZS) FluxInputsDynamicStepr image_latentsimage_latent_inputsadditional_batch_inputsct|ts|g}t|ts|g}||_||_t|y)N) isinstancelist_image_latent_inputs_additional_batch_inputssuper__init__)rrGrH __class__s rrOzFluxInputsDynamicStep.__init__sH -t4#6"7 148'>&? #$7!(?% rrcd}d}|js |jr>d}|jr|d|jz }|jr|d|jz }d}||z|zS)NzInput processing step that: 1. For image latent inputs: Updates height/width if None, patchifies latents, and expands batch size 2. For additional batch inputs: Expands batch dimensions to match final batch sizez Configured inputs:z - Image latent inputs: z - Additional batch inputs: zN This block should be placed after the encoder steps and the text input step.)rLrM)rsummary_section inputs_infoplacement_sections rrz!FluxInputsDynamicStep.descriptions c   $ $(E(E2K((!*C*CAB*G 'M<"."<"<#[%;%;=P=V=VWX=Y[hjv#  #>20&1&G&G&11 #  K!8:M N= OB77 ;J"; ;L#7%)&1&G&G&11 L K\ : ; UK05  r)r=r>r?r@rrBrOrArr r&r r+rr r< __classcell__)rPs@rrErE}sJ+:):-/ !#Y "&c ASAA,Z($ d;&7  5!#65!}5!Q^5!rrEc$eZdZdZdededefdZy)FluxKontextInputsDynamicStep flux-kontextr0r3rc|j|}|jD]}t||}|t||j\}}t |ds||_t |ds||_|jdd\}} tj||j|jd|| }t|||j|j}t||||jD]A} t|| } | t| | |j|j} t|| | C|j!||||fSrd)r6rLrfr rgrhr_rar.rrir(r rrjrMr9rks rr<z%FluxKontextInputsDynamicStep.__call__s{**51 (,'@'@ O #")+7N"O "*==PR\RmRmnMFE;7+1 (; 6*/ '+>*C*CAB*G 'M<"."<"<#[%;%;=P=V=VWX=Y[hjv#  #>20&1&G&G&11 #  K!8:M N9 O>77 ;J"; ;L#7%)&1&G&G&11 L K\ : ; UK05  rN)r=r>r?r@rr r<rrrrqrqs#J3!#63!}3!Q^3!rrqcreZdZdZdZedeefdZedee fdZ e dZ de dedefd Zy ) FluxKontextSetResolutionSteprrc y)NzDetermines the height and width to be used during the subsequent computations. It should always be placed _before_ the latent preparation step.rrs rrz(FluxKontextSetResolutionStep.description3s  O rrcXtdtdtdtdg}|S)NrXrYrZmax_areai)rWr"r)r r*)rr&s rr&z#FluxKontextSetResolutionStep.inputs9s. H % G $ J#w G   rcJtdtdtdtdgS)NrXz'The height of the initial noisy latentsr`rZz&The width of the initial noisy latentsrbrs rr+z1FluxKontextSetResolutionStep.intermediate_outputsBs( XBk l WAi j  rc|||dzzdk7rtd|dzd|| ||dzzdk7rtd|dzd|yy)NrrzHeight must be divisible by z but is zWidth must be divisible by )r/)rXrZrgs rr2z)FluxKontextSetResolutionStep.check_inputsIs  &,!?1!D:;Ka;O:PPXY^X_`a a"E rr0r3c|j|}|jxs |j}|jxs |j}|j |||j ||}}|j}||z } t|| zdz}t|| z dz}|j dz} || z| z}|| z| z}||k7s||k7rtjd|d|d||_||_|j||||fS)Ng?rz6Generation `height` and `width` have been adjusted to z and z to fit the model requirements.) r6rXdefault_heightrZ default_widthr2rgrxroundloggerwarningr9) rr0r3r1rXrZoriginal_heightoriginal_widthrx aspect_ratio multiple_ofs rr<z%FluxKontextSetResolutionStep.__call__Qs'**51 ##@z'@'@!!=Z%=%= &%)D)DE*0%''v~ x,.367</C78 11A5 ${2;&4 _ $(? NNHPUV[U\\{| $ !  UK05  rN)r=r>r?r@rrArr r&r r+ staticmethodr2rr r<rrrruru0s~J Z( d;&7  bb!#6!}!Q^!rru)typingrr$ pipelinesrutilsrmodular_pipelinerr modular_pipeline_utilsr r qwenimage.inputsr r r get_loggerr=rrrErqrurrrrsq %C<]1   H %Y!-Y!zw!1w!t6!#86!r;!#8;!r