bi\ddlmZddlmZmZmZmZddlZddlm cm Z ddl m Z ddlmZddlmZmZddlmZmZmZmZmZmZmZmZd d lmZer dd lmZm Z m!Z!m"Z"dd l#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*ejVe,Z-Gd dZ.GddZ/GddZ0GddZ1y))Path)DictListOptionalUnionN)validate_hf_hub_args) safe_open)_LOW_CPU_MEM_USAGE_DEFAULTload_state_dict)USE_PEFT_BACKEND_get_detailed_type_get_model_file_is_valid_typeis_accelerate_availableis_torch_versionis_transformers_availablelogging)_maybe_expand_lora_scales)CLIPImageProcessorCLIPVisionModelWithProjectionSiglipImageProcessorSiglipVisionModel) AttnProcessorAttnProcessor2_0IPAdapterAttnProcessorIPAdapterAttnProcessor2_0IPAdapterXFormersAttnProcessorJointAttnProcessor2_0!SD3IPAdapterJointAttnProcessor2_0c eZdZdZe d deeeeeee jffdeeeefdeeeefde efdZ dZ dZy ) IPAdapterMixinMixin for handling IP Adapters.%pretrained_model_name_or_path_or_dict subfolder weight_nameimage_encoder_folderc t|ts|g}t|ts|g}t|dk(r|t|z}t|ts|g}t|dk(r|t|z}t|t|k7r tdt|t|k7r td|j dd}|j dd}|j dd}|j d d} |j d d} |j d d} |j d t } | r!t sd} tjd | durtdds tdddd} g}t|||D]\}}}t|tst|||||| | | ||  }|jdriid}t|dd5}|j!D]t}|j#dr(|j%||d|j'dd<<|j#dsN|j%||d|j'dd<v dddnt)|}n|}t|j!}d|vrd|vr td |j+|t-|d!rt/|d!d|t|tstj1d"||j3d#d$k(rt5||j7}nt5|j7}t9j:||| || |j<%j?|j@}|jC|&n td'tjd(t-|d)s8t/|d)dGd*}|jD |jDjFjHn|}tK||+}|jC|,t-|d-st/||jLn |jN}|jQ|| .|jS|}|ik7rxtTstjd/yt/|d0i}|jWD];\}}d1||vs|jY|d1|2|j[d1|gd3g4=yy#1swYxYw)5a Parameters: pretrained_model_name_or_path_or_dict (`str` or `List[str]` or `os.PathLike` or `List[os.PathLike]` or `dict` or `List[dict]`): Can be either: - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on the Hub. - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved with [`ModelMixin.save_pretrained`]. - A [torch state dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict). subfolder (`str` or `List[str]`): The subfolder location of a model file within a larger model repository on the Hub or locally. If a list is passed, it should have the same length as `weight_name`. weight_name (`str` or `List[str]`): The name of the weight file to load. If a list is passed, it should have the same length as `subfolder`. image_encoder_folder (`str`, *optional*, defaults to `image_encoder`): The subfolder location of the image encoder within a larger model repository on the Hub or locally. Pass `None` to not load the image encoder. If the image encoder is located in a folder inside `subfolder`, you only need to pass the name of the folder that contains image encoder weights, e.g. `image_encoder_folder="image_encoder"`. If the image encoder is located in a folder other than `subfolder`, you should pass the path to the folder that contains image encoder weights, for example, `image_encoder_folder="different_subfolder/image_encoder"`. cache_dir (`Union[str, os.PathLike]`, *optional*): Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used. force_download (`bool`, *optional*, defaults to `False`): Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist. proxies (`Dict[str, str]`, *optional*): A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. local_files_only (`bool`, *optional*, defaults to `False`): Whether to only load local model weights and configuration files or not. If set to `True`, the model won't be downloaded from the Hub. token (`str` or *bool*, *optional*): The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from `diffusers-cli login` (stored in `~/.huggingface`) is used. revision (`str`, *optional*, defaults to `"main"`): The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git. low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): Speed up model loading only loading the pretrained weights and not initializing the weights. This also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this argument to `True` will raise an error. rT`weight_name` and `pretrained_model_name_or_path_or_dict` must have the same length.8`weight_name` and `subfolder` must have the same length. cache_dirNforce_downloadFproxieslocal_files_onlytokenrevisionlow_cpu_mem_usage,Cannot initialize model with low cpu memory usage because `accelerate` was not found in the environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install `accelerate` for faster and less memory-intense model loading. You can do so with: ``` pip install accelerate ``` .T>=1.9.0~Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set `low_cpu_mem_usage=False`.attn_procs_weightspytorch file_type framework weights_namer,r-r.r/r0r1r& user_agent .safetensors image_proj ip_adapterptcpur;device image_proj.rA ip_adapter.rBNRequired keys are (`image_proj` and `ip_adapter`) missing from the state dict. image_encoderloading image_encoder from /rr&r2r,r/ torch_dtyperKa`image_encoder` cannot be loaded because `pretrained_model_name_or_path_or_dict` is a state dict.image_encoder is not loaded since `image_encoder_folder=None` passed. You will not be able to use `ip_adapter_image` when calling the pipeline with IP-Adapter.Use `ip_adapter_image_embeds` to pass pre-generated image embedding instead.feature_extractorsize crop_sizerSunetr2/PEFT backend is required to load these weights. peft_configfaceid_ adapter_name?adapter_weights). isinstancelistlen ValueErrorpopr rloggerwarningrNotImplementedErrorzipdictrendswithr keys startswith get_tensorreplacer appendhasattrgetattrinfocountras_posixrfrom_pretraineddtypetorFregister_modulesrKconfig image_sizer unet_namerY_load_ip_adapter_weights_load_ip_adapter_lorasr itemsload_lora_weights set_adapters)selfr%r&r'r(kwargsr,r-r.r/r0r1r2r> state_dicts model_file state_dictfkeyrnimage_encoder_subfolderrKdefault_clip_sizeclip_image_sizerSrY extra_lorasr\kloras W/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/loaders/ip_adapter.pyload_ip_adapterzIPAdapterMixin.load_ip_adapter9s"x+t,&-K?F5Z4[ 1 4 5 :4Y\_`k\l4l 1)T*" I y>Q !C $44I { s#HI Ist t { s9~ -WX XJJ{D1 $4e<**Y-!::&8$? 7D)::j$/"JJ': %  NN.   $-=dG-L%.  $8iP  MP 1; N H K I 1; CTJ,9!,'#1#%5%')  ''702"!EJ":eLmPQ#$668mC"~~m<[\[g[ghk[l < 8]TV9W X!$ !>[\[g[ghk[l < 8]TV9W X mmm"1! " #PQ&dM2> *002RGAt }K7..tGA3-.P))WQC=/C5)QR {mmsAS3:)S33S= c t|dst||jn |j}t |t s|g}t ||d}|jjD]\}}t |tttfs#t|t|jk7r.tdt|dt|jdt|dk(r|t|jz}t|D]^\}}t |t r:|jD]&\}} |j#|s| |j|<(P||j|<`y)  Set IP-Adapter scales per-transformer block. Input `scale` could be a single config or a list of configs for granular control over each IP-Adapter behavior. A config can be a float or a dictionary. Example: ```py # To use original IP-Adapter scale = 1.0 pipeline.set_ip_adapter_scale(scale) # To use style block only scale = { "up": {"block_0": [0.0, 1.0, 0.0]}, } pipeline.set_ip_adapter_scale(scale) # To use style+layout blocks scale = { "down": {"block_2": [0.0, 1.0]}, "up": {"block_0": [0.0, 1.0, 0.0]}, } pipeline.set_ip_adapter_scale(scale) # To use style and layout from 2 reference images scales = [{"down": {"block_2": [0.0, 1.0]}}, {"up": {"block_0": [0.0, 1.0, 0.0]}}] pipeline.set_ip_adapter_scale(scales) ``` rY default_scaleCannot assign  scale_configs to IP-Adapter.rN)rsrtr~rYrcrdrattn_processorsrrrrrescalerf enumeraterlro) rrrY scale_configs attn_nameattn_processori scale_configrss rset_ip_adapter_scalez#IPAdapterMixin.set_ip_adapter_scales\<5>> # Assuming `pipeline` is already loaded with the IP Adapter weights. >>> pipeline.unload_ip_adapter() >>> ... ``` rKNrPsafety_checkerrSrXtext_encoder_hid_proj text_projscaled_dot_product_attention)rsrtrKregister_to_configrSrYencoder_hid_projr|encoder_hid_dim_typerrrFrrrcrrr __class__set_attn_processorr attn_procsnamevalueattn_processor_classs runload_ip_adapterz IPAdapterMixin.unload_ip_adapter1s 4 )gdOT.R.^!%D   # #4, # ?t-.t01gdDWY]6^6j)-&''4,'G&* "04 - 4995 6499;Z;Z;f)-)H)HDII &.2DII +4?DII   1 9944::< KD%&-a1O&P "VcVe ! 24MOmn%__& t   $$Z0rNrP)__name__ __module__ __qualname____doc__rrstrrrtorchTensorrrrrrrr#r#6s) /> @R/4S$s)T#u||J[E\5\/]@Rd3i(@R3S >* @R 'sm @R@RD3?j/1rr#c eZdZdZedeeeeeee jffdeeeefdeeeeffdZ dZ dZ y) ModularIPAdapterMixinr$r%r&r'c t|ts|g}t|ts|g}t|dk(r|t|z}t|ts|g}t|dk(r|t|z}t|t|k7r tdt|t|k7r td|j dd}|j dd}|j dd}|j d d}|j d d} |j d d} |j d t } | r!t sd} tjd | durtdds tdddd} g} t|||D]-\}}}t|tst||||||| | ||  }|jdriid}t|dd5}|j!D]t}|j#dr(|j%||d|j'dd<<|j#dsN|j%||d|j'dd<v dddnt)|}n|}t|j!}d|vrd|vr td | j+|0t-|d!d"}t-||}|j/| | #|j1| }|ik7rxt2stjd$yt-|d%i}|j5D];\}}d&||vs|j7|d&|'|j9d&|gd(g)=yy#1swYxYw)*a Parameters: pretrained_model_name_or_path_or_dict (`str` or `List[str]` or `os.PathLike` or `List[os.PathLike]` or `dict` or `List[dict]`): Can be either: - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on the Hub. - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved with [`ModelMixin.save_pretrained`]. - A [torch state dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict). subfolder (`str` or `List[str]`): The subfolder location of a model file within a larger model repository on the Hub or locally. If a list is passed, it should have the same length as `weight_name`. weight_name (`str` or `List[str]`): The name of the weight file to load. If a list is passed, it should have the same length as `subfolder`. cache_dir (`Union[str, os.PathLike]`, *optional*): Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used. force_download (`bool`, *optional*, defaults to `False`): Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist. proxies (`Dict[str, str]`, *optional*): A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. local_files_only (`bool`, *optional*, defaults to `False`): Whether to only load local model weights and configuration files or not. If set to `True`, the model won't be downloaded from the Hub. token (`str` or *bool*, *optional*): The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from `diffusers-cli login` (stored in `~/.huggingface`) is used. revision (`str`, *optional*, defaults to `"main"`): The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git. low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): Speed up model loading only loading the pretrained weights and not initializing the weights. This also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this argument to `True` will raise an error. rr*r+r,Nr-Fr.r/r0r1r2r3Tr4r5r6r7r8r9r<r?r@rCrDrErGrArHrIrBrJr~rYrZr[r\r]r^r`ra)rcrdrerfrgr rrhrirrjrkrlrrmr rnrorprqr rrrtrrr rrr)rr%r&r'rr,r-r.r/r0r1r2r>rrrrrrnr~rYrr\rrs rrz%ModularIPAdapterMixin.load_ip_adapterfsh+t,&-K?F5Z4[ 1 4 5 :4Y\_`k\l4l 1)T*" I y>Q !C $44I { s#HI Ist t { s9~ -WX XJJ{D1 $4e<**Y-!::&8$? 7D)::j$/"JJ': %  NN.   $-=dG-L%.  ."  MP 1; N ! + I 1; CTJ,9!,'#1#%5%')  ''702"!EJ":eLmPQ#$668mC"~~m<[\[g[ghk[l < 8]TV9W X!$ !>[\[g[ghk[l < 8]TV9W X mmm"1! " #PQ&dM2> *002RGAt }K7..tGA3-.P))WQC=/C5)QR -mmsAM6:)M66N c t|dd}t||}t|ts|g}t||d}|jj D]\}}t|t ttfs#t|t|jk7r.tdt|dt|jdt|dk(r|t|jz}t|D]^\}}t|tr:|j D]&\} } |j| s| |j|<(P||j|<`y ) rr~rYrrrrrrN)rtrcrdrrrrrrrerrfrrlro) rrr~rYrrrrrrrs rrz*ModularIPAdapterMixin.set_ip_adapter_scalesW<D+v6 tY'%&GE1$SQ )-)=)=)C)C)E ? %I~!79RTr s}%^-A-A)BB$(]);(<* TRTRl4?l'1rrceZdZdZedddej fdeee ee eejffdeee efde eee efde ede ed ejf d Zd eee ee e effd Zd Zy)FluxIPAdapterMixinz$Mixin for handling Flux IP Adapters.rHrKr%r'r&+image_encoder_pretrained_model_name_or_pathrimage_encoder_dtypec t|ts|g}t|ts|g}t|dk(r|t|z}t|ts|g}t|dk(r|t|z}t|t|k7r tdt|t|k7r td|j dd}|j dd} |j dd} |j d d} |j d d} |j d d} |j d t }|r!t sd}tjd |durtdds tdddd}g}t|||D]\}}}t|ts= 1.9.0 else `False`): Speed up model loading only loading the pretrained weights and not initializing the weights. This also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this argument to `True` will raise an error. rr*r+r,Nr-Fr.r/r0r1r2r3Tr4r5r6r7r8r9r<r?r@rCrDrEzip_adapter_proj_model.rGzdouble_blocks.rIc3@K|]}j|ywNro.0prefixrs r z5FluxIPAdapterMixin.load_ip_adapter..s"Xf3>> "X.rAc3@K|]}j|ywrrrs rrz5FluxIPAdapterMixin.load_ip_adapter..s$ZS^^F%;$Zrip_adapter_double_stream_k_projto_k_ipip_adapter_double_stream_v_projto_v_ipz processor.rHrBrJrKrLrNrPrQrRrSrTrUrXrZ)'rcrdrerfrgr rrhrirrjrkrlrrmr rnanyjoinsplitrprqr rrrsrtrurrxrzrFevalr{rKr|r}r transformerr)rr%r'r&rrrrr,r-r.r/r0r1r2r>rrrrimage_proj_keysip_adapter_keysdiffusers_namernrKrrrSrs @rrz"FluxIPAdapterMixin.load_ip_adapter`s|+t,&-K?F5Z4[ 1 4 5 :4Y\_`k\l4l 1)T*" I y>Q !C $44I { s#HI Ist t { s9~ -WX XJJ{D1 $4e<**Y-!::&8$? 7D)::j$/"JJ': %  NN.   $-=dG-L%.  $8iP  MP 1; N P K I 1; CTJ,9!,'#1#%5%')  ''702"!EJ":eL]PQ+C]*S+;]*K#$668 ]C""X"XX14#))C.:L1MKL< !/ LM<J%&KTR &ABmAn$op9II K*A2C*31A,?  R _!TV&--M-J(NNg t01gdDWY]6^6f$'!<@.Ks#JjD6Is1v#Js,,zExpected list of z scales, got z, c32K|]}t|ywr)r)rxs rrz:FluxIPAdapterMixin.set_ip_adapter_scale..NsFkRSs1vFksrN)rintfloatrrnum_ip_adaptersr|rrcrangerr TypeErrorrrerfrrdrrkrvaluesr) rr scale_typer_rinvalid_scale_sizesrrrs @rrz'FluxIPAdapterMixin.set_ip_adapter_scale s43:& **;;KK%%,,77  eZ ($)/$:;qU;E; E4 #3 4A9MGEtE*d:>N2N,O'PQ./A%/H.IUV V u: (~c%j\_DUUbcd d JUJ J38"PaJq$*}dE#tCy.12 }d 6>c] }d "*# }d#[[}d}d~6)%tE{DeDU0U*V6)p&8rrc eZdZdZedefdZe d dee e e e jffde de e de e ddf d Zd eddfd Zdd Zy)SD3IPAdapterMixinz1Mixin for handling StableDiffusion 3 IP Adapters.returnc|jjjDcgc]}t|tr |j  }}t |dkDxrtd|DScc}w)a7Checks if IP-Adapter is loaded and scale > 0. IP-Adapter scale controls the influence of the image prompt versus text prompt. When this value is set to 0, the image context is irrelevant. Returns: `bool`: True when IP-Adapter is loaded and any layer has scale > 0. rc3&K|] }|dkD yw)rNr)rrs rrz9SD3IPAdapterMixin.is_ip_adapter_active..s&EUuqy&Es)rrrrcr!rrer)r attn_procscaless ris_ip_adapter_activez&SD3IPAdapterMixin.is_ip_adapter_activesj"--==DDF )%FG OO  6{QE3&Ef&E#EE  s#A.Nr%r'r&r(c 2|jdd}|jdd}|jdd}|jdd} |jdd} |jdd} |jd t} | r!tsd} tj d | d urt d d s t dddd} t|tst|||||| | | ||  }|jdriid}t|dd5}|jD]t}|jdr(|j||d|jdd<<|jdsN|j||d|jdd<v dddnt!|}n|}t#|j}d|vrd|vr t%dt'|drt)|dd|t|tstj+d||j-d d!k(rt/||j1}nt/|j1}| || d"}|j3t5j6|fi|t9j6|fd#|j:i|j=|j>$n t%d%tj d&|j@jC|| 'y#1swYgxYw)(a Parameters: pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`): Can be either: - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on the Hub. - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved with [`ModelMixin.save_pretrained`]. - A [torch state dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict). weight_name (`str`, defaults to "ip-adapter.safetensors"): The name of the weight file to load. If a list is passed, it should have the same length as `subfolder`. subfolder (`str`, *optional*): The subfolder location of a model file within a larger model repository on the Hub or locally. If a list is passed, it should have the same length as `weight_name`. image_encoder_folder (`str`, *optional*, defaults to `image_encoder`): The subfolder location of the image encoder within a larger model repository on the Hub or locally. Pass `None` to not load the image encoder. If the image encoder is located in a folder inside `subfolder`, you only need to pass the name of the folder that contains image encoder weights, e.g. `image_encoder_folder="image_encoder"`. If the image encoder is located in a folder other than `subfolder`, you should pass the path to the folder that contains image encoder weights, for example, `image_encoder_folder="different_subfolder/image_encoder"`. cache_dir (`Union[str, os.PathLike]`, *optional*): Path to a directory where a downloaded pretrained model configuration is cached if the standard cache is not used. force_download (`bool`, *optional*, defaults to `False`): Whether or not to force the (re-)download of the model weights and configuration files, overriding the cached versions if they exist. proxies (`Dict[str, str]`, *optional*): A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. local_files_only (`bool`, *optional*, defaults to `False`): Whether to only load local model weights and configuration files or not. If set to `True`, the model won't be downloaded from the Hub. token (`str` or *bool*, *optional*): The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from `diffusers-cli login` (stored in `~/.huggingface`) is used. revision (`str`, *optional*, defaults to `"main"`): The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier allowed by Git. low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): Speed up model loading only loading the pretrained weights and not initializing the weights. This also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this argument to `True` will raise an error. r,Nr-Fr.r/r0r1r2r3Tr4r5r6r7r8r9r<r?r@rCrDrErGrArHrIrBrJrKrLrMr)r2r,r/rO)rSrKrQrRrZ)"rgr rrhrirrjrcrlrrmr rnrorprqr rdrfrsrtrurvrrwr{rrxrryrzrFrr)rr%r'r&r(rr,r-r.r/r0r1r2r>rrrrrnrs rrz!SD3IPAdapterMixin.load_ip_adapters2rJJ{D1 $4e<**Y-!::&8$? 7D)::j$/"JJ': %  NN.   $-=dG-L%.  $8iP ?F(5(#-!1!#% J##N3,.bA zT%HiA vvxi>>-8WXWcWcdgWhJ|4S[[PR5ST ^^M:WXWcWcdgWhJ|4S[[PR5ST iii-Z8 >JJOO%& t # D(@mn n 4 )gdOT.R.Z#/!"GNKK"=>c=d ef+11#6!;26yBV2W2`2`2b/267K2L2U2U2W/.?%.,<F ))*>*N*NOf*qjp*q&7&G&G3'AE'OU'"T[[/ *%{c 11*Pa1bciisAL .)L  Lrc|jjjD]}t|ts||_y)am Set IP-Adapter scale, which controls image prompt conditioning. A value of 1.0 means the model is only conditioned on the image prompt, and 0.0 only conditioned by the text prompt. Lowering this value encourages the model to produce more diverse images, but they may not be as aligned with the image prompt. Example: ```python >>> # Assuming `pipeline` is already loaded with the IP Adapter weights. >>> pipeline.set_ip_adapter_scale(0.6) >>> ... ``` Args: scale (float): IP-Adapter scale to be set. N)rrrrcr!r)rrrs rrz&SD3IPAdapterMixin.set_ip_adapter_scale*s;&#..>>EEG -N.*KL',$ -rc t|dr&t|ddd|_|jdt|dr&t|ddd|_|jdd|j _|j jjDcic]0\}}|t|tr tn|j2}}}|j j|ycc}}w)z Unloads the IP Adapter weights. Example: ```python >>> # Assuming `pipeline` is already loaded with the IP Adapter weights. >>> pipeline.unload_ip_adapter() >>> ... ``` rKNrPrSrX)rsrtrKrrSrrArrrcr!r rr)rrrrs rrz#SD3IPAdapterMixin.unload_ip_adapterAs 4 )gdOT.R.^!%D   # #$ # 7 4, -'$@SUY2Z2f%)D "  # #d # ;'+# $//??EEG  e +5e=^+_%'ejetetev w  ++J7  s5C1)zip-adapter.safetensorsNrK)rN)rrrrpropertyboolrrrrrrrrrrrrrrrrrs; FdFF"4#'.= Qc/4S$sELL?P:Q5Q/RQcQcC= Qc 'sm Qc QcQcf-%-D-. 8rr)2pathlibrtypingrrrrrtorch.nn.functionalnn functionalrhuggingface_hub.utilsr safetensorsr models.modeling_utilsr r utilsr rrrrrrrunet_loader_utilsr transformersrrrrmodels.attention_processorrrrrrr r! get_loggerrrhr#rrrrrrr s.. 6!O   9ww   H %j1j1Z w1w1ta8a8H `8`8r