biw ddlmZmZmZddlZddlmZddlmZddl m Z ddl m Z m Z mZmZdd lmZmZmZd d l mZe j*eZGd d eZGddeZGdde ZGddeZy))AnyListTupleN)FluxTransformer2DModel)FlowMatchEulerDiscreteScheduler)logging) BlockStateLoopSequentialPipelineBlocksModularPipelineBlocks PipelineState) ComponentSpec InputParam OutputParam)FluxModularPipelinec eZdZdZedeefdZedefdZ edee ee ffdZ e jdededed e j$def d Zy ) FluxLoopDenoiserfluxreturnc$tdtgS)N transformer)rrselfs c/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/modular_pipelines/flux/denoise.pyexpected_componentsz$FluxLoopDenoiser.expected_components&sm-CDEEc y)NzStep within the denoising loop that denoise the latents. This block should be used to compose the `sub_blocks` attribute of a `LoopSequentialPipelineBlocks` object (e.g. `FluxDenoiseLoopWrapper`)rs r descriptionzFluxLoopDenoiser.description*  5 rc jtdtddtjdtddtjdtddtjd td dtjd td dtjd tddtjdgS)Njoint_attention_kwargslatentsTz^The initial latents to use for the denoising process. Can be generated in prepare_latent step.required type_hintr!guidancezGuidance scale as a tensor prompt_embedszPrompt embeddingspooled_prompt_embedszPooled prompt embeddingstext_idsz/IDs computed from text sequence needed for RoPElatent_image_idsz0IDs computed from image sequence needed for RoPE)rtorchTensorrs rinputszFluxLoopDenoiser.inputs2s / 0 ,,|   ,,8   ,,/   &,,6   ,,M   ",,N  A' ' r components block_stateitc |j|j|jdz |j|j|j |j |j|jd d}||_ ||fS)NiF) hidden_statestimestepr)encoder_hidden_statespooled_projectionsr$txt_idsimg_ids return_dictr) rr%flattenr)r*r+r$r,r- noise_pred)rr1r2r3r4r>s r__call__zFluxLoopDenoiser.__call__]s ++%--YY[4' ))"-";";*??#.#E#E((00,     ", ;&&rN)__name__ __module__ __qualname__ model_namepropertyrrrstrr!rrr0r.no_gradrr intr/rr?r rrrr#sJ FT-%8FF S  ( U38_-( ( TU]]_'-'to)rr1r2r3r4 latents_dtypes rr?zFluxLoopAfterDenoiser.__call__s$++11 (2277  " "    8      $ $ 5"-"5"5"8"8"GK ;&&rN)r@rArBrCrDrrrrEr!rrr0rPrrRr.rFrr rGr/r?r rrrIrIqsJ MT-%8MM S  U38_-)T#Y))dd;&7ddU]]_ '#6 'Z 'TW '\a\h\h ' 'rrIceZdZdZedefdZedeefdZ edee fdZ e jdededefdZy ) FluxDenoiseLoopWrapperrrc y)NzPipeline block that iteratively denoise the latents over `timesteps`. The specific steps with each iteration can be customized with `sub_blocks` attributesr rs rr!z"FluxDenoiseLoopWrapper.descriptions  d rcBtdttdtgS)NrKr)rrrrs rloop_expected_componentsz/FluxDenoiseLoopWrapper.loop_expected_componentss$ +'F G -)? @  rcbtddtjdtddtdgS)N timestepsTzWThe timesteps to use for the denoising process. Can be generated in set_timesteps step.r&num_inference_stepszgThe number of inference steps to use for the denoising process. Can be generated in set_timesteps step.)rr.r/rGrs r loop_inputsz"FluxDenoiseLoopWrapper.loop_inputss? ,,u   %F    rr1statec|j|}tt|j|j|j j zz d|_|j jd|j|j5}t|jD]{\}}|j||||\}}|t|jdz k(s3|dz|jkDsL|dz|j j zdk(sl|j} ddd|j||||fS#1swYxYw)Nr)total)r3r4r)get_block_statemaxlenr^r_rKordernum_warmup_stepsset_begin_index progress_bar enumerate loop_stepupdateset_block_state)rr1rar2rjr3r4s rr?zFluxDenoiseLoopWrapper.__call__s;**51 '*  %% &)H)H:K_K_KeKe)e egh( $ ,,Q/   [%D%D  E *!+"7"78 *1*...[TUYZ.*[' KK112Q66Uk:::AI]I]IcIc?cgh?h '')  * * UK05   * *sA!E3EEE N)r@rArBrCrDrEr!rrr\rr`r.rFrrr?r rrrYrYsJ  S   $}*=   T*-  U]]_!#6!}!Q^!!rrYc2eZdZeegZddgZedefdZ y)FluxDenoiseStepdenoiserafter_denoiserrc y)Na*Denoise step that iteratively denoise the latents. Its loop logic is defined in `FluxDenoiseLoopWrapper.__call__` method At each iteration, it runs blocks defined in `sub_blocks` sequencially: - `FluxLoopDenoiser` - `FluxLoopAfterDenoiser` This block supports both text2image and img2img tasks.r rs rr!zFluxDenoiseStep.descriptions  E rN) r@rArBrrI block_classes block_namesrDrEr!r rrrprps1%'<=M/0K  S  rrp)typingrrrr.modelsr schedulersrutilsr modular_pipeliner r r rmodular_pipeline_utilsrrrr get_loggerr@loggerrrIrYrpr rrr~s|$# ,9 LK1   H %K',K'\)'1)'X6!96!r  ,  r