a Ad,@snddlZddlZddlZddlZddlmZGdddZGdddeZ GdddeZ Gd d d eZ dS) N)sharedc@s$eZdZddZddZddZdS) ReplaceHelpercCs g|_dSN)replacedselfrI/var/www/html/stable-diffusion-webui/modules/sd_disable_initialization.py__init__ szReplaceHelper.__init__cCs:t||d}|durdS|j|||ft||||Srgetattrrappendsetattrrobjfieldfuncoriginalrrr replace s   zReplaceHelper.replacecCs,|jD]\}}}t|||q|jdSr)rrclear)rrrrrrr restoreszReplaceHelper.restoreN)__name__ __module__ __qualname__r rrrrrr r s rcs:eZdZdZd fdd ZddZddZd d ZZS) DisableInitializationa When an object of this class enters a `with` block, it starts: - preventing torch's layer initialization functions from working - changes CLIP and OpenCLIP to not download model weights - changes CLIP to not make requests to check if there is a new version of a file you already have When it leaves the block, it reverts everything to how it was before. Use it like this: ``` with DisableInitialization(): do_things() ``` Tcst||_dSr)superr disable_clip)rr __class__rr r .s zDisableInitialization.__init__cCs:t||d}|durdS|j|||ft||||Srr rrrr r2s   zDisableInitialization.replacecs,dd}ddfdd }fdd}fdd }d d d d fdd }d d fdd }d d fdd }tjjd|tjjd|tjjd|jr(td|_tjj jj d|_ t j jd|_t jd|_t jd|_t jjd|_dS)Nc_sdSrrargskwargsrrr do_nothing=sz3DisableInitialization.__enter__..do_nothing) pretrainedcsj|ddi|S)Nr#)create_model_and_transforms)r#r r!rrr .create_model_and_transforms_without_pretrained@szWDisableInitialization.__enter__..create_model_and_transforms_without_pretrainedcs(jdg|R|id|}||_|S)N)config state_dict)CLIPTextModel_from_pretrained name_or_path)pretrained_model_name_or_path model_argsr!resrrr r(CszFDisableInitialization.__enter__..CLIPTextModel_from_pretrainedcs,|ddd|dd}j|i|S)Nr)/)1transformers_modeling_utils_load_pretrained_modelrrrr r0HszZDisableInitialization.__enter__..transformers_modeling_utils_load_pretrained_modelc_s|dks|dkr |ddkr dSzB||g|Rddi|}|dur^||g|Rddi|}|WSty||g|Rddi|YS0dS)NzShttps://huggingface.co/openai/clip-vit-large-patch14/resolve/main/added_tokens.jsonzopenai/clip-vit-large-patch14rzadded_tokens.jsonlocal_files_onlyTF) Exception)rurlr r!r,rrr *transformers_utils_hub_get_file_from_cacheLs zSDisableInitialization.__enter__..transformers_utils_hub_get_file_from_cacheF)r1csj|g|Ri|Sr)%transformers_utils_hub_get_from_cacher3r1r r!rr4rr r5ZszNDisableInitialization.__enter__..transformers_utils_hub_get_from_cachecsj|g|Ri|Sr)0transformers_tokenization_utils_base_cached_filer6r7rr r8]szYDisableInitialization.__enter__..transformers_tokenization_utils_base_cached_filecsj|g|Ri|Sr),transformers_configuration_utils_cached_filer6r7rr r9`szUDisableInitialization.__enter__..transformers_configuration_utils_cached_filekaiming_uniform__no_grad_normal__no_grad_uniform_r$from_pretrained_load_pretrained_model cached_fileZget_from_cache)rtorchnninitr open_clipr$ldmmodulesencoders CLIPTextModelr( transformersmodeling_utilsPreTrainedModelr0tokenization_utils_baser8configuration_utilsr9utilshubr5)rr"r%r(r0r5r8r9rr7r __enter__<s$  zDisableInitialization.__enter__cCs |dSrrrexc_typeexc_valexc_tbrrr __exit__oszDisableInitialization.__exit__)T) rrr__doc__r rrOrU __classcell__rrrr rs  3rc@s eZdZdZddZddZdS)InitializeOnMetaa Context manager that causes all parameters for linear/conv2d/mha layers to be allocated on meta device, which results in those parameters having no values and taking no memory. model.to() will be broken and will need to be repaired by using LoadStateDictOnMeta below when loading params from state dict. Usage: ``` with sd_disable_initialization.InitializeOnMeta(): sd_model = instantiate_from_config(sd_config.model) ``` cstjjr dSdd|tjjdfdd|tjjdfdd|tjjdfdd|tjj dd ddS) NcSs d|d<|S)Nmetadevicer)xrrr set_devicesz.InitializeOnMeta.__enter__..set_devicer cs|i|Srrr) linear_initr\rr z,InitializeOnMeta.__enter__..cs|i|Srrr) conv2d_initr\rr r^r_cs|i|Srrr)mha_initr\rr r^r_toc_sdSrrrrrr r^r_) rcmd_opts&disable_model_loading_ram_optimizationrr@rALinearConv2dMultiheadAttentionModulerr)r`r]rar\r rOszInitializeOnMeta.__enter__cCs |dSrrPrQrrr rUszInitializeOnMeta.__exit__N)rrrrVrOrUrrrr rXss  rXcs:eZdZdZd fdd ZddZddZd d ZZS) LoadStateDictOnMetaa Context manager that allows to read parameters from state_dict into a model that has some of its parameters in the meta device. As those parameters are read from state_dict, they will be deleted from it, so by the end state_dict will be mostly empty, to save memory. Meant to be used together with InitializeOnMeta above. Usage: ``` with sd_disable_initialization.LoadStateDictOnMeta(state_dict): model.load_state_dict(state_dict, strict=False) ``` Ncs2t||_||_|pi|_|jd|_dS)N)rr r'rZweight_dtype_conversionget default_dtype)rr'rZrkrrr r s   zLoadStateDictOnMeta.__init__cCs |dd\}}|j||jS)N.)splitrkrlrm)rkeyZkey_first_term_rrr get_weight_dtypesz$LoadStateDictOnMeta.get_weight_dtypecstjjr dS j j fddd fdd  tjjd fdd tjjd fd d tjj d fd d tjj d fd d tjj d fd d tjj d fdd tjj d fdddS)Nc s g}|jD]\}}|dur q||} | d} | dur\| j| d|| <|| |jr| durp| jn|j} tj j j tj || d|j d|j|<q|jD]2}||} | d} | dur| || <|| q||||g|Ri||D]} || dqdS)N)dtyperZrt) requires_grad) _parametersitemspoprbrsr is_metartr@rA parameter Parameter zeros_likerv_buffers) rmoduler'prefixr r!Zused_param_keysnameparamrqZsd_paramrt)rZsdrrr load_from_state_dicts*  (   z;LoadStateDictOnMeta.__enter__..load_from_state_dictTcs,|krdd|D}||||ddS)atorch makes a lot of copies of the dictionary with weights, so just deleting entries from state_dict does not help because the same values are stored in multiple copies of the dict. The trick used here is to give torch a dict with all weights on meta device, i.e. deleted, and then it doesn't matter how many copies torch makes. In _load_from_state_dict, the correct weight will be obtained from a single dict with the right weights (sd). The dangerous thing about this is if _load_from_state_dict is not called, (if some exotic module overloads the function and does not call the original) the state dict will just fail to load because weights would be on the meta device. cSs"i|]\}}||jd|jdqS)rYru)rbrt).0kvrrr r_zJLoadStateDictOnMeta.__enter__..load_state_dict..)strictN)rx)rrr'r)rrr load_state_dicts z6LoadStateDictOnMeta.__enter__..load_state_dictrcsg|Ri|Srrr)rmodule_load_state_dictrr r^r_z/LoadStateDictOnMeta.__enter__.._load_from_state_dictcsg|Ri|Srrr)rmodule_load_from_state_dictrr r^r_csg|Ri|Srrr)linear_load_from_state_dictrrr r^r_csg|Ri|Srrr)conv2d_load_from_state_dictrrr r^r_csg|Ri|Srrr)rmha_load_from_state_dictrr r^r_csg|Ri|Srrr)layer_norm_load_from_state_dictrrr r^r_csg|Ri|Srrr)group_norm_load_from_state_dictrrr r^r_)T)rrcrdr'rZrr@rArhrerfrg LayerNorm GroupNormrr) rrZrrrrrrrrrrr rOszLoadStateDictOnMeta.__enter__cCs |dSrrPrQrrr rUszLoadStateDictOnMeta.__exit__)N) rrrrVr rsrOrUrWrrrr ris  >ri) ldm.modules.encoders.modulesrDrCr@Ztransformers.utils.hubrHrErrrrXrirrrr s U