^j׆DddlZddlZddlZddlmZddlmZmZmZm Z ddl m Z ddl m Z ddlmZddlmZmZmZmZmZmZmZmZmZdd lmZdd lmZerddlZerdd lm Z dd l!m"Z"m#Z#d Z$ejJe&Z'erddl(m)Z)m*Z*GddZ+defdZ,y)N)replace) TYPE_CHECKINGAnyLiteralOptional) safe_open)PeftConfigLike)get_model_conversion_mapping) CONFIG_NAME cached_filecheck_peft_versionextract_commit_hashfind_adapter_config_fileis_accelerate_availableis_peft_availableis_torch_availablelogging)DownloadKwargs)log_state_dict_report)dispatch_model)get_balanced_memoryinfer_auto_device_mapz0.19.1)LoadStateDictConfigLoadStateDictInfoceZdZUdZdZdZedzed<eee fed< d'dedzdedzdeee fdzd eed fdzd e d e d e e dzde deee fdzde dddfdZ d(dede dddfdZd)dedzddfdZdeeezddfdZd*dZd*dZdeefdZd+dedzdedzdefdZ d,d ed!edzd"edzd#edzddf d$Zd%eeezddfd&Zy)-PeftAdapterMixinu' A class containing all functions for loading and using adapters weights that are supported in PEFT library. For more details about adapters and injecting them on a transformer-based model, check out the documentation of PEFT library: https://huggingface.co/docs/peft/index Currently supported PEFT methods are all non-prompt learning methods (LoRA, IA³, etc.). Other PEFT models such as prompt tuning, prompt learning are out of scope as these adapters are not "injectable" into a torch module. For using these methods, please refer to the usage guide of PEFT library. With this mixin, if the correct PEFT version is installed (>= 0.19.1), it is possible to: - Load an adapter stored on a local path or in a remote Hub repository, and inject it in the model - Attach new adapters in the model and train them with Trainer or by your own. - Attach multiple adapters and iteratively activate / deactivate them - Activate / deactivate all adapters from the model. - Get the `state_dict` of the active adapter. FN_prepare_peft_hotswap_kwargs peft_configauto peft_model_id adapter_nameadapter_state_dictz torch.Tensorlow_cpu_mem_usage is_trainablehotswaplocal_files_onlyadapter_kwargs load_configrreturnrc ./ddlm.ddlm} ddlm} m}m}|rd| d<| | jni}|j| |jd d| d0i|} |xs | j}|d k(r8t|d d }t|jxr||jv}|xr|}|r\|jr||jvr t!d t#.fd|jj%Dr t!d||nd}| xsi} ddlm}m}|jr|s||jvrt!d|d|r%|jr||jvr t!d || | t!d|`| j*jd0i| t-|fi| j*}|t!d|d|j.|fi| j*}ddlm}t5|}t7|dr| |_||||}|s |||||h/|@t|dd3/j;|j<j>jAdtBdtf/fd }|jsd|_||ddg}| jDd ur|jGdx}}d}|D]>} ||dd|jId r | jDnd dd || j*!\}}n||xs tKd"gi}}t|d#d$|jLi} d }!|jOD]}"t|"d%d}#|#d}!n|!rtQ}$||}%n|h|djId rT|Ri}%|D]J}&tS|&d&d'(}'|$j;|'|'jUD]}(|'jW|(|%|(<Ln.|!i}%|D]})|%j||)n t!d)|%}t#d*|j%Dr t!d+| |||tY| |||| ,} |j[|||| |j]D*+cgc]\}*}+||*s|*c}+}*-\},}+|j8rIdd.l/m0}-|jc|jOD] }"te|"|-s|"jgd "|,jhD(chc] }(||(s |(c}(|,_4tk|| j| jl|,tn/|,S#tJ$r }|}Yd}~d}~wwxYwcc}+}*wcc}(w)1u Load adapter weights from file or remote Hub folder. If you are not familiar with adapters and PEFT methods, we invite you to read more about them on PEFT official documentation: https://huggingface.co/docs/peft Requires PEFT to be installed as a backend to load the adapter weights. Args: peft_model_id (`str`, *optional*): The identifier of the model to look for on the Hub, or a local path to the saved adapter config file and adapter weights. adapter_name (`str`, *optional*): The adapter name to use. If not set, will use the name "default". load_config (`LoadStateDictConfig`, *optional*): A load configuration to reuse when pulling adapter weights, typically from `from_pretrained`. kwargs (`dict[str, Any]`, *optional*): Additional `LoadStateDictConfig` fields passed as keyword arguments. peft_config (`dict[str, Any]`, *optional*): The configuration of the adapter to add, supported adapters are all non-prompt learning configs (LoRA, IA³, etc). This argument is used in case users directly pass PEFT state dicts. adapter_state_dict (`dict[str, torch.Tensor]`, *optional*): The state dict of the adapter to load. This argument is used in case users directly pass PEFT state dicts. low_cpu_mem_usage (`bool`, *optional*, defaults to `False`): Reduce memory usage while loading the PEFT adapter. This should also speed up the loading process. is_trainable (`bool`, *optional*, defaults to `False`): Whether the adapter should be trainable or not. If `False`, the adapter will be frozen and can only be used for inference. hotswap : (`"auto"` or `bool`, *optional*, defaults to `"auto"`) Whether to substitute an existing (LoRA) adapter with the newly loaded adapter in-place. This means that, instead of loading an additional adapter, this will take the existing adapter weights and replace them with the weights of the new adapter. This can be faster and more memory efficient. However, the main advantage of hotswapping is that when the model is compiled with torch.compile, loading the new adapter does not require recompilation of the model. When using hotswapping, the passed `adapter_name` should be the name of an already loaded adapter. If the new adapter and the old adapter have different ranks and/or LoRA alphas (i.e. scaling), you need to call an additional method before loading the adapter: ```py model = AutoModel.from_pretrained(...) max_rank = ... # the highest rank among all LoRAs that you want to load # call *before* compiling and loading the LoRA adapter model.enable_peft_hotswap(target_rank=max_rank) model.load_adapter(file_name_1, adapter_name="default") # optionally compile the model now model = torch.compile(model, ...) output_1 = model(...) # now you can hotswap the 2nd adapter, use the same name as for the 1st # hotswap is activated by default since enable_peft_hotswap was called model.load_adapter(file_name_2, adapter_name="default") output_2 = model(...) ``` By default, hotswap is disabled and requires passing `hotswap=True`. If you called `enable_peft_hotswap` first, it is enabled. You can still manually disable it in that case by passing `hotswap=False`. Note that hotswapping comes with a couple of limitations documented here: https://huggingface.co/docs/peft/main/en/package_reference/hotswap adapter_kwargs (`dict[str, Any]`, *optional*): Additional keyword arguments passed along to the `from_pretrained` method of the adapter config and `find_adapter_config_file` method. r)PeftType)_maybe_shard_state_dict_for_tpr )r_get_resolved_checkpoint_filesload_state_dictTr'Npretrained_model_name_or_pathr _hotswap_enabledFz\To hotswap an adapter, there must already be an existing adapter with the same adapter name.c3PK|]}|jjk7ywN) peft_typeLORA).0confr,s i/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/integrations/peft.py z0PeftAdapterMixin.load_adapter..sYt4>>X]]2Ys#&zMHotswapping is currently only supported for LoRA, please set `hotswap=False`.default PeftConfiginject_adapter_in_modelAdapter with name - already exists. Please use a different name.zhYou should either pass a `peft_model_id` or a `peft_config` and `adapter_state_dict` to load an adapter.z adapter model file not found in zB. Make sure you are passing the correct path to the adapter model.)build_peft_weight_mappinginference_mode)rr4keyr*c.tfdDS)Nc3&K|]}|v ywr3)r6markerrBs r8r9zHPeftAdapterMixin.load_adapter..is_adapter_key..sGv}Gs)any)rBadapter_key_markerss`r8is_adapter_keyz5PeftAdapterMixin.load_adapter..is_adapter_keysG3FGG Gzadapter_model.safetensorszadapter_model.binz .safetensors)r0variant gguf_fileuse_safetensors user_agentis_remote_codetransformers_explicit_filenamedownload_kwargsz@Could not download either a .bin or a .safetensors adapter file. hf_device_map_tp_infoptcpu) frameworkdevicez5Neither a state dict nor checkpoint files were found.c3RK|]}t|tj !ywr3) isinstancetorchTensor)r6vs r8r9z0PeftAdapterMixin.load_adapter..6sXqz!U\\22Xs%'z3J;//PR'##$CTJ)=,<= %R)R)R f %d,>FO $T%@%@%glVZVfVfFf h %;*;G //\IYIY5Y rYt?O?O?V?V?XYY !pqq'3'?|Y '-2<  & &ldN^N^>^1,?lmn n t:: TXTdTd@dn   &8&@[EXz    .K ' ' . . @ @":#--#  #* 6}oF%% 5*44--K X9$? ; 0 1-9)9K &#<Q R **e3!))+26 6 /J$5 ' '9W6C $"&;K;T;TUc;dK77jo#'',7G(3(C(C :6$&6 '$ ' oG,n$oo13R. T?R4EF  lln Ffj$7G""&   %K!-$6! ,$Q'00@&.$&!,JD#,TT%#PLOOL1)..0J/;/F/Fq/I)!,JJ "-$&!!1II%,,_Y-GHI!!XYY!2 X)>)@VANSTDU1V6  a  % % ? IIK,,. 1fn5))%0 11=0I0I$_1^\]M^Q$_ !*5*S*S$/$G$G%  i'!&J'@W%`s05:T$T=-T= U.U$ T:-T55T: target_rankcheck_compiled)rwarnignorect|dir?|dk(r td|dk(rtjdn|dk7rt d|dd |_||d |_y ) aEnables the possibility to hotswap PEFT adapters with different ranks, or, if the model is compiled, without triggering recompilation. Right now, hotswapping is only supported for LoRA. Calling this method is only required when hotswapping adapters and if the model is compiled or if the ranks of the loaded adapters differ. If the ranks are all identical and the model is not compiled, hotswapping works without calling this method first. Args: target_rank (`int`, *optional*, defaults to `128`): The highest rank among all the adapters that will be loaded. check_compiled (`str`, *optional*, defaults to `"error"`): How to handle the case when the model is already compiled, which should generally be avoided. The options are: - "error" (default): raise an error - "warn": issue a warning - "ignore": do nothing rrzr?z8adapter_config should be an instance of PeftConfig. Got z instead. name_or_pathN)rMIN_PEFT_VERSIONrjr<r=rrrrsrZ TypeErrortypermgetbase_model_name_or_path set_adapter)radapter_configr"r<r=s r8 add_adapterzPeftAdapterMixin.add_adapters$ '78<#0y ***.D ' T-- -1,?lmn n.*5VW[\jWkVlluvw w261B1B>SW1X.lC &rJc |tt|js tdt |t rpt |t |jz }t|dkDrtddj|dt |jj||jvr2td|dt |jjdd l m }dd l m}d }|jD]'\}}t |||fs|j!|d })|s td y)a If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT official documentation: https://huggingface.co/docs/peft Sets a specific adapter by forcing the model to use a that adapter and disable the other adapters. Args: adapter_name (`Union[list[str], str]`): The name of the adapter to set. Can be also a list of strings to set multiple adapters. r0No adapter loaded. Please load an adapter first.rz)Following adapter(s) could not be found: , zV. Make sure you are passing the correct adapter name(s). current loaded adapters are: r>z7 not found. Please pass the correct adapter name among reModulesToSaveWrapperFTzhDid not succeeded in setting the adapter. Please make sure you are using a model that supports adapters.N)rrrrrsrZlistrrlenjoinrrrf peft.utilsr named_modulesr)rr"missingrfr_adapters_has_been_setrrs r8rzPeftAdapterMixin.set_adaptersZ '78**OP P  d +,'#d.>.>*??G7|a ? '@R?ST559$:J:J:O:O:Q5R4SU!1!1 1$\N2ijnosoopEpEpGkHjIJ  <3!&++- .IAv&>3G"HI""<0)-& . &z &rJctt|js tdddlm}ddlm}|jD]&\}}t|||fs|jd(y) a, If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT official documentation: https://huggingface.co/docs/peft Disable all adapters that are attached to the model. This leads to inferring with the base model only. rrrrerFenabledN) rrrrrsrrfrrrrZenable_adapters)rrfrrrs r8disable_adaptersz!PeftAdapterMixin.disable_adaptersse '78**OP P;3++- 6IAv&>3G"HI&&u&5 6rJctt|js tdddlm}|j D]$\}}t||s|jd&y)z If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT official documentation: https://huggingface.co/docs/peft Enable adapters that are attached to the model. rrrreTrN) rrrrrsrrfrrZr)rrfrrs r8rz PeftAdapterMixin.enable_adapterss\ '78**OP P;++- 5IAv&.1&&t&4 5rJctt|js tdddlm}|j D]\}}t||s|j}nttr|g}|S)a3 If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT official documentation: https://huggingface.co/docs/peft Gets the current active adapters of the model. In case of multi-adapter inference (combining multiple adapters for inference) returns the list of all active adapters so that users can deal with them accordingly. For previous PEFT versions (that does not support multi-adapter inference), `module.active_adapter` will return a single string. rrrre) rrrrrsrrfrrZactive_adapterr{)rrfrractive_adapterss r8rz PeftAdapterMixin.active_adaptersss '78**OP P;++- IAv&.1"("7"7  os +./OrJrbctt|js tdddlm}||j d}||||}|S)ax If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT official documentation: https://huggingface.co/docs/peft Gets the adapter state dict that should only contain the weights tensors of the specified adapter_name adapter. If no adapter_name is passed, the active adapter is used. Args: adapter_name (`str`, *optional*): The name of the adapter to get the state dict from. If no name is passed, the active adapter is used. state_dict (nested dictionary of `torch.Tensor`, *optional*) The state dictionary of the model. Will default to `self.state_dict()`, but can be used if special precautions need to be taken when recovering the state dictionary of a model (like when using model parallelism). rrr)get_peft_model_state_dict)rbr")rrrrrsrjrr)rr"rbrr#s r8get_adapter_state_dictz'PeftAdapterMixin.get_adapter_state_dictsT '78**OP P2  //1!4L6t amn!!rJr` max_memoryoffload_folder offload_indexci}dtjtjvr||d<|j}|dk7rt ||||dk(}t |trt|||}t|f||d|y)a Optional re-dispatch the model and attach new hooks to the model in case the model has been loaded with accelerate (i.e. with `device_map=xxx`) Args: device_map (`str` or `dict[str, Union[int, str, torch.device]]` or `int` or `torch.device`, *optional*): A map that specifies where each submodule should go. It doesn't need to be refined to each parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the same device. If we only pass the device (*e.g.*, `"cpu"`, `"cuda:1"`, `"mps"`, or a GPU ordinal rank like `1`) on which the model will be allocated, the device map will map the entire model to this device. Passing `device_map = 0` means put the whole model on GPU 0. To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For more information about each option see [designing a device map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map). max_memory (`Dict`, *optional*): A dictionary device identifier to maximum memory. Will default to the maximum memory available for each GPU and the available CPU RAM if unset. offload_folder (`str` or `os.PathLike`, *optional*): If the `device_map` contains any value `"disk"`, the folder where we will offload weights. offload_index (`int`, *optional*): The offload_index argument to be passed to `accelerate.dispatch_model` method. r sequentialbalanced_low_0)rno_split_module_classeslow_zero)rr)r` offload_dirN) inspect signaturer parameters_no_split_modulesrrZr{r)rr`rrrdispatch_model_kwargsrs r8_dispatch_accelerate_modelz+PeftAdapterMixin._dispatch_accelerate_model5s<!# g//?JJ J5B !/ 2"&"8"8  %,%(?$(88 J j# &.E\J   !& $  rJ adapter_namesctt|js tdddlm}t |tr|g}|Dcgc]}||jvs|}}|rtddj||Dcgc]6}|j|jjjd8}}t||D]G\}}||||t|d d st|d s,|jj!|d It#|jdk(r |`d |_y y cc}wcc}w) z Delete a PEFT adapter from the underlying model. Args: adapter_names (`Union[list[str], str]`): The name(s) of the adapter(s) to delete. rrr)delete_adapterz@The following adapter(s) are not present and cannot be deleted: rr)r"prefixrrFrN)rrrrrspeft.functionalrrZr{rrr4ryrzziprprwpopr)rrrnamemissing_adaptersr"prefixesrs r8rzPeftAdapterMixin.delete_adaptermsI '78**OP P2 mS )*OM.;d$JZJZ>ZD[[ RSWS\S\]mSnRop fssUat'' 5??EEKKMNaPss$' x$@ 9 L& 4l6 Jt5u='$P]B^  $$\48  9 t A % */D ' &\ tsD8D8;D=) NNNNFFr FNN)rr3)r*N)NN)NNN)__name__ __module__ __qualname____doc__rrrdict__annotations__r{r rrqrrrintrrrrrrrrrrrErJr8rr9s$#04 $+4c>)**%)#'-1?C"'"*0!&047;NTzNDjN#s(^d* N !n!45< N  NN'NNS#X-N34N Nb\c#k#k6=>W6X#k #kJ&'d &'d&'P'S C'D'R6&5$c:"3:"RVY]R]"im"@"&%)$( 6 6 $J6 d 6 Tz 6  6 p%0DIO%0%0rJrrQc v| tsd||fS|jd}|jdt|t|jdt |jdd|jdt |jdd||jd|jd ddd }t |d|d<|j d d}|j dd}||j}d D]}||vs||vs ||||<d|vr|d|d <t |jdd|d<t |jdd|d<|xs||d<t|fi|}|tjj|rt|dd5} |}tjj|rAtjjtjj|tstj | d}ddd|||fS#1swYxYw)Ntoken commit_hash cache_dirforce_downloadFproxiesr'revision subfolder) rrrr'rrr _raise_exceptions_for_gated_repo%_raise_exceptions_for_missing_entries'_raise_exceptions_for_connection_errors_adapter_model_path)rrr _commit_hashrzutf-8)encodingr)rrr r rqrrcopyrospathisfileopenexistsrjsonload) r0rQr(rresolved_config_filertoken_from_adapter_kwargs peft_kwargsarg_namefs r8maybe_load_adaptersr s^ %,4E4G2NBB    (E=)1* ) %))+6 3 34De LM#'' 2!/"5"56H%"PQ$((4%))+6-22749 *==QSW)X &(,,-BDI . 2 27D A"$))+ = BH +(o2M(7(A H% B O +*9-*HK '(,_-@-@AQSX-Y(Z $%*./B/BCUW\/]*^ &'$A(A G6 )   &277>>:M+N %sW = X"?  77>>"?@ :KHI15 ! =V0W- X  =~ MM X Xs 'A;H//H8)-rrr dataclassesrtypingrrrr safetensorsr_typingr conversion_mappingr utilsr r rrrrrrr utils.hubrutils.loading_reportrr[ accelerateraccelerate.utilsrrr get_loggerrrirlrrrr rErJr8rs 88!$=   '8)K   H %GY 0Y 0x:N#:NrJ