oiGfSSKJr SSKrSSKrSSKrSSKrSSKJr SSKrSSK r SSKJ r J r SSK J r Jr SSKJr SSKJr SSKJr S S KJr S S KJrJrJrJrJrJrJr S S K J!r! S r"Sr#SSjr$SSSjjr%SSjr&SSSjjr'SS.Sjr(SSSjjr)g)) annotationsN)Optional) file_existshf_hub_download)EntryNotFoundErrorLocalEntryNotFoundError) load_file)http_user_agent)PEFT_TYPE_TO_PREFIX_MAPPING)INCLUDE_LINEAR_LAYERS_SHORTHAND)EMBEDDING_LAYER_NAMESSAFETENSORS_WEIGHTS_NAME WEIGHTS_NAMEAuxiliaryTrainingWrappercheck_file_exists_on_hf_hub infer_devicematch_target_against_key)PeftTypec[US5=(aH [UR[RR [RR 45$)z.Check if the layer has an embedding base layer base_layer)hasattr isinstancertorchnnLinear Embedding)layers R/home/james-whalen/.local/lib/python3.13/site-packages/peft/utils/save_and_load.pyhas_valid_embedding_base_layerr ,s; 5, ' oJu7G7G%((//[`[c[c[m[mIn,oocxUR5H&up4U(dXA:XdU[USS5:XdM$Us $ g)z7Get the name of the embedding module for a given layer.rN) named_modulesgetattr)modelris_embedding_in_target_modulesnamemodules rget_embedding_layer_namer)1s<++- .6?vQXY^`lnrQsGsK. r!c ^^!^"^#U(a [USU5nURTm!UcUR5nT!R[R [R 4;GaT!RnUS:XaUVs0sHnSU;dM XaU_M nnOtUS:Xa$UVs0sHnSU;dSU;dMXaU_M nnOJUS:Xa>0nUH5nSU;dM XXv'URS5SS-nX;dM/XXx'M7 O[eUR5VV s0sHupiSU;aTU;dSU;dMXi_M nnn T!R[R :Xa`T!Rn U bQU R5VV s0sHupiURS T3S 5U _M n nn U T!l URXT5nT!R(a7S TS 3m"U"4S jn UR5VV s0sHupiU "U5U _M nnn GOT!R[R:XaT!RnUS:XaUVs0sHnSU;dM XaU_M nnGOXUS:Xa%UVs0sHnSU;dSU;dMXaU_M nnGO-US:Xa?0nUH5nSU;dM XXv'URS5SS-nX;dM/XXx'M7 GO[eT!R[R :Xa@UVs0sH1ofRS 5SR#S5(dM,XaU_M3 nnGOT!R$(a0nT!R[R&:Xa\UR(TR*US'UR(TR,US'UR(TR.R0n OFT!R2(a$UR(TR.R0n OUR5T5n XS'GOT!R[R6:Xa[8T!Rn UVs0sHomU;dM XaU_M nn[:R<"5S:Xa[>R@"S5 URC5H|up[EUS5(dMURFR5HFupi[:R<"5S:XaU RI[JRL5OU X~SU3'MH M~ GOT!R[RN:Xas[8T!RnUVs0sHnUU;dM XaU_M nnT!RP(a0ST3U;a [SS5eUST-UST-'UST-UST-'GO T!R[RT:XaUVs0sHnSU;dM XaU_M nnGOT!R[RV:XGaJ0nT!RXS:a[JRZnORT!RXS:a[JR\nO1T!RXS:a[JR^nO[JR`nT!Rb(aUHnS U;dM XReT!Rd5unnURgUS!-URIUS"905 URgUS#-[JRh"USS$9SS2SS2SS24Rk505 M OUVs0sHnS U;dM XaU_M nnUS%T-US%T-'OfT!R[m[5;a1[8T!RnUVs0sHnUU;dM XaU_M nnO[SS&T!R35eURC5Hup[oU[p5(dMUR#S'5(aURsS'5nUR5VV s0sH5upiUR#US 35(dM URsUS 35U _M7 nnn URgURuTU5R5VV s0sH upiUS U3U _M sn n5 M S(n[ET!S)5(a[oT!Rv[x5(a[T!Rv[z:waG[EUS*5(aUR}5OUn[U!4S+jURC555nO*T!Rv(a[U!4S,j[55nT!R[R:H=(d [T!S-S5SLnUS.:Xa(U(a!U(d[>R@"S/5 S0nGOUS.:XGa[[US1S5S2S5n[T!S3S5nS(nUbt[RR[RRUS455nU=(d [US45nUc[>R@"S5US635 S(nOUnU(aZU(aSU(aLUURRRU5R:wa[>R@"S75 S0nOS(nU(a[EUS85(aUR5UR54HnnU(a[U5(dM[UUU5nU(dM2URgUR5VV s0sHupiUU;dM Xi_M sn n5 Mp OU(a[>R@"S95 [R"[R"S T35S:-5m#UU!U#4S;jn UR5VV s0sHupiU "U5U _M nnn U$s snfs snfs sn nfs sn nfs sn nfs snfs snfs snfs snfs snfs snfs snfs snfs sn nfs sn nfs sn nfs sn nf)>URT5(aUSSnU$)Niendswith)knew_dora_suffixs rrenamed_dora_weights7get_peft_model_state_dict..renamed_dora_weightss"::o..#2Ar!boft_ boft_only adaption_prefix_task_colsprefix_task_rowsprompt_embeddingsWindowszWindows has issues saving integers into safetensors. Hence, we convert shira_indices to float32 before saving on Windows OS. The shira_indices will always be converted to integers when loading. shira_indices.shira_indices.zbase_model.vera_A.zModel was initialised to not save vera_A and vera_B but config now specifies to save projection! Set `config.save_projection` to `False`.zbase_model.vera_B.internal_xlora_classifieril vblora_logits _topk_indices)dtype _topk_weightsdimzbase_model.vblora_vector_bank.zUnknown PEFT type passed: _fsdp_wrapped_module.Ftarget_modulesget_base_modelc3>^# UH>umn[U4Sj[55(dM&[TRT5v M@ g7f)c3Z># UH n[R"SUS3T5v M" g7f)z(.*\.)?$N)rematch).0er8s r 6get_peft_model_state_dict... s)S=Rrxx71#Q33=Rs(+N)anyrrrO)rV_r8configs @rrX,get_peft_model_state_dict.. s?(2DAqS=RSSC()>)>BB2s %A A c3@># UHoTR;v M g7f)N)rO)rVr8r\s rrXr]s'bLaqV-B-B(BLastrainable_token_indicesautozXSetting `save_embedding_layers` to `True` as embedding layers found in `target_modules`.Tr\ vocab_sizebase_model_name_or_pathz config.jsonz Could not find a config file in z4 - will assume that the vocabulary was not modified.zdSetting `save_embedding_layers` to `True` as the embedding layer has been resized during finetuning.get_input_embeddingsuYCould not identify embedding layer(s) because the model is not a 🤗 transformers model.rScj>SU;aU$URST35(aURST35$URS5upnTR[R :Xa3UR TS35(aUS-URTS35-$TRSU5nUSU3$)Nr1r[r2) r7 removesuffix rpartition peft_typerVBLORA startswith removeprefixsub)keyr[suffix adapter_namer\patterns rremove_adapter_name6get_peft_model_state_dict..remove_adapter_nameKs c>J <* + +##a ~$67 7,    /V5F5F,WXGY5Z5Z9v22l^13EFF Fkk"c"ax  r!)Rr$ peft_config state_dictrgrLORAADALORAr/splitNotImplementedErroritems rank_patternreplace!resize_state_dict_by_rank_patternuse_doraBOFTADAPTION_PROMPTriis_prompt_learningMULTITASK_PROMPT_TUNINGprompt_encoderr@rA embeddingweightinference_modeget_prompt_embedding_to_saveSHIRAr platformsystemwarningswarnr#rrDtorfloat32VERAsave_projection ValueErrorXLORArh num_vectorsuint8int16int32int64save_only_topk_weightstopkupdatesoftmax contiguouslistrrrjadapter_state_dictrOstrr rPrZrTRAINABLE_TOKENSospathexistsjoinrr\ __class__from_pretrainedrarcget_output_embeddingsr r)rTcompileescape)$r%rsrnunwrap_compiledsave_embedding_layersr/r8 to_return bias_namevryr:rB shira_prefixr'r( vera_prefix indices_dtypelogitsindicesprefixmodule_state_dictembedding_is_targeted_modelusing_trainable_tokensramodel_idhas_base_configlocal_config_existsrrembedding_module_namerpr\r9ros$ ` @@@rget_peft_model_state_dictr9s :{E2   | ,F%%' HMM8+;+;<<{{ 6>3=N:aA)qM):INI U]3=]:aAQW[\Q\)qM):I]I [ Ia<#-=IL ! 0 3f 3=N:aA)qM):INI U]3=]:aAQW[\Q\)qM):I]I [ Ia<#-=IL ! 0 3f 2*D @>HH\_kHk=lI*\9 :=GH\_kHk=lI*\9 :   X^^ +/9^z!=X\]=]%Q1 %z ^   X__ ,    $!KKM   % '!KKM   % '!KKM!KKM  ( ("a'&0m&8&8&EOFG$$a/&97::M:;Z%[\$$a/&95==UW;XYZ\]_b`b_bYb;c;n;n;p%qr  4>V:aTUAU)qM):IVEO ,| ;F 2\AB   T(^ +,V-=-=>/9Iz!Vq[%Q1 %z I 5f6F6F5GHII++-  f6 7 7677(()@A ;E:J:J:L!:L$!PQP\P\`d_eef]gPh-$qz*A-:L !   .4.G.G Vg.h.n.n.pq.pdaD61#!.pq .:"v'(( f++S 1 1v7L7LPo7o07u>N/O/OU))+UZF$'("002(% !  " "$''bLa'b$b ! H555uIbdh9iqu9u&+@I_ pq $ & (WUHd;\4P 6#G!*8A7HI7Htq$Q'*7HII {O]t gS O]g"P&O_(W J!r\&i:Js o5 o5$o:8 o:!o?=o?"p4p  p p'p; p5+p$ p? p  p  p% p% p* p*' p/5 p/ p4 p4p9>p9p?  q q q cU(dU/4$/nUR5nUR5HupVXT;aM XERSS:Xa)XER5S-UR5:XaMJXERUR:wdMhUR XVRXER45 M UH un nX M X4$)Nr>r )rsrxshapenumelappend)r%peft_model_state_dictignore_mismatched_sizes mismatchedrsrltensorr[s r_find_mismatched_keysrds #$b((J!!#J,224    O ! !" % *1F1F1H11LPVP\P\P^1^  ? FLL 0   sLL*/2G2GH I5 Q ! &  ! ,,r!cL0nUR5HupEX$;aURU5u pgSU;aXSRURS5SS5n[R "[R "U5S-USU3U5nOUSU3nXSU'MXSU'M U$)zbUtility function to remap the state_dict keys to fit the PEFT model by inserting the adapter name.r1r NrS)rxrfrrvrTrkr) rsrnparameter_prefixrrlvalr[rmsuffix_to_replaces r$_insert_adapter_name_into_state_dictrs$$&  ">>*:;LAqf}$'HHV\\#->qr-B$C!ffRYY'89D@\NRSTeSfBgilmQ|n-),# &),# &' ! r!c ^ URUnUnUR5Huupx[U[5(dMUR U5n UR S5(aUR S5nU Hn USU 3n USX3n XXl'Xk M Mw UR(dUR[R:XaUnGOUR[R:XaUnGOUR[;Ga0n[URn UR[R:XGa`UR(GaNURUR up[#UR%55nUGHn SU ;dM XjR'[(R*5nU R-SS5nXjR-SS5n[(R."USUR1SSS 9- /SS 9n[(R2"U5n[(R4"/UR S SQUP5R7[9S 55R'UR:5R=SUU5nUUU'Xj XjR-SS5 GM [?XbU S 9nUR[R@:Xa#URBnUbUREUU5 GOUR[RF:Xa[HRJ"5S:Xa[LRN"S5 UR5Hgupx[QUS5(dMUSU3U;dM%URSUSU35nUR'[(RT5URVU'Mi GO3UR[RX:XaxURZ(aSU;a []S5eURZ(dSU;a[LRN"S5 OURZ(d[LRN"S5 OUR[R^:Xa8SU3m U 4SjnURa5V Vs0sHun nU"U 5U_M nn nOGUR[Rb:Xa"[eSU55(a []S5eO[fe[iXUS9unnU(aMURkUSSS9nURm5H'n[QUS5(dMURoU5 M) OURkUSS9nUR(a,URpURrRkSUS 0SS9 UR[Rt:XaURpURkUSS9 U(agS!RwUVVVs/sHunnnS"US#US$US%3PM snnn5nS&URxRzS'US3n[LRN"U5 U$s snn fs snnnf)(a@ Set the state dict of the PEFT model. Given a PEFT `state_dict` (as returned by [`get_peft_model_state_dict`]), insert the weights into the model. The model needs to have the PEFT adapters already in place (e.g. via [`inject_adapter_in_model`]). Setting the adapter weights also takes care of re-inserting the adapter name. This name may be a different name than the one originally used to train the adapter. Args: model ([`PeftModel`]): The Peft model. peft_model_state_dict (`dict`): The state dict of the Peft model. adapter_name (`str`, *optional*, defaults to `"default"`): The name of the adapter whose state dict should be set. ignore_mismatched_sizes (`bool`, *optional*, defaults to `False`): Whether to ignore mismatched in the state dict. low_cpu_mem_usage (`bool`, `optional`, defaults to `False`): This argument must be `True` if the `model` was loaded with adapter weights on the meta device, e.g. after calling `inject_adapter_in_model` with `low_cpu_mem_usage=True`. Otherwise, leave it as `False`. rNr1rIr2rKr r>T)keepdimrLNz-inf)rnrrCzWindows has issues saving integers into safetensors. Hence, we had converted shira_indices to float32 before saving on Windows OS. The shira_indices will always be converted to integers when loading.rDrEzbase_model.vera_AzXSpecified to load vera_A and vera_B from state dictionary however they were not present!zSpecified to not load vera_A and vera_B from state dictionary however they are present in state dictionary! Consider using them to ensure checkpoint loading is correct on all platforms using `peft_config.save_projection = True`zSpecified to not load vera_A and vera_B from state dictionary. This means we will be relying on PRNG initialisation to restore these projections using `config.projection_prng_key`, which may not be accurate on all system configurations.r3c>>URT5(aUS-nU$)Nr4r6)r8old_dora_suffixs rr:7set_peft_model_state_dict..renamed_dora_weightss ::o..I Ar!c3,# UH nSU;v M g7f)z.oft_r.N)rVrls rrX,set_peft_model_state_dict..$sE/D9#/DszTrying to load old OFT checkpoint, which is no longer supported. Please install PEFT <= v0.15.2 to load it or train a new OFT adapter.)rF)strictassign%_move_adapter_to_device_of_base_layer)rrrB z- z: found shape z in the checkpoint and z in the model instantiatedzSome weights of zy were not initialized from the model checkpoint and are being ignored because you passed `ignore_mismatched_sizes=True`: )>rrr#rradapter_state_dict_load_maprirjrrgrr~rr rhrvblora_vector_bankrrkeysrrlongrzcatsumlogzerosfill_floatdevicescatterrruryresize_modules_by_rank_patternrrrrrrpopintrDrrrrtrxOFTrZrwrload_state_dictmodulesrrrrrr__name__)!r%rrnrlow_cpu_mem_usager\rsr'r(key_mapr8 lookup_key store_keyrrr[state_dict_keysr original_key topk_weights topk_logitsmatrixryshira_indices_valuesr:mismatched_keys load_resultrlshape1shape2mismatched_warningmsgrs! @rset_peft_model_state_dictrs<  | ,F&J ++-  f6 7 788FG677(()@A $vQqc] #fAgj\2 (=(I %*.&  F$4$48P8P$P *   X^^ + *   8 8 "6v7G7GH   x .63P3P3P"55lCIINK":??#45O$#a'" ((4A#$99_b#AL#-ii.Y#ZL#(99lA @P@PQS]a@P@bR=T=TUZU^U^=_,,\:!6   .%%*=EZ*Z n++0CG\0\ < ++ E    .!7|nEO  MbLgLgLi$jLiDAq%9!%U.*?++,A%X\+] mmoFvFGG<<\J&++,A%+P     \*44DD ,-@A B4 E 8;;; \*::;PY^:_!YY,; +:'CSEx/FvhNhi+:  u7789XXjWkkl n   c U%k@ s X;#Y T) weights_onlyc2[R"USU0UD6$)z|Call torch.load and handle weights_only. Defaults to weights_only=True to anticipate upcoming switch on the PyTorch side. r)rload)rargskwargss r torch_loadrPs ::t A, A& AAr!c .^TRSS5b#[RRUTS5OUnUc [ 5nSU4SjjnST;a [ 5TS'[RR [RRU[55(a([RRU[5nSnGO=[RR [RRU[55(a'[RRU[5nSnO[RR(a*U"SS9nTRSS5 [X4SS0TD6nSnOTRS S5n U cTRS S5n U"SS9n[UUTRS S5TRS S5U S 9n U nU (a[U[40TD6nO[U[40TD6nU(aN[%[&R(S5(a%U[&R*"S5:Xa [-USS9n O([-XaS9n O[/U[&R*"U5S9n U(dU n U $0n U R15HupU R3S5(aSnO$U R3S5(aSnO [#S5eU R5U5n UR15H+unn[6R8"UUU 5unnUS:dM)Un O UU 3nXU'M U $![a U"SS9n[X4SS0TD6nSnGNUf=f![ a$ [#SUSUS[S[SUS3 5ef=f)a A helper method to load the PEFT weights from the HuggingFace Hub or locally Args: model_id (`str`): The local path to the adapter weights or the name of the adapter to load from the HuggingFace Hub. device (`str`): The device to load the weights onto. key_mapping (dict, *optional*, defaults to None) Extra mapping of PEFT `state_dict` keys applied before loading the `state_dict`. When this mapping is applied, the PEFT-specific `"base_model.model"` prefix is removed beforehand and the adapter name (e.g. `"default"`) is not inserted yet. Only pass this argument if you know what you're doing. hf_hub_download_kwargs (`dict`): Additional arguments to pass to the `hf_hub_download` method when loading from the HuggingFace Hub. subfolderNTc>U(a[O[nTRSS5b#[RR TSU5$U$)Nr)rrgetrrr)use_safetensors weights_namehf_hub_download_kwargss rget_hub_filename+load_peft_weights..get_hub_filenametsK3B/ &))+t<H GGLL/ /@|, ww~~bggll4)ABCC77<<&>?  T<8 9 977<<l3  " " 1 1'= ""#5t< $&xoPToXnoH"O'**7D9 =*../?FE'= &1!+// DA,00dC ' #6 &&()H  *8\\E[\ 5>>5 ) )ve9L/L-huE -hF %hU\\&=QR $4!4 %$-%'!(..0HC~~122, ..& w ""6*C(3(9(9(;$%'WWWk3%G"q=!C )< "(.O9&  -hZtH:F22>tD\C]]lmulvvwy s5L?<M&? M#"M#&.N)NdefaultFr`)F)r%ztorch.nn.Modulerdict[str, torch.Tensor]rboolreturnzRtuple[dict[str, torch.Tensor], list[tuple[str, tuple[int, ...], tuple[int, ...]]]])rsr&rnrrrr(r&)r%FF)rr'rr'r(None)NN)rrrz Optional[str]rzOptional[dict[str, str]]r(dict)* __future__rrrrTrtypingrrrrrhuggingface_hub.errorsrrsafetensors.torchr rtransformers.utilsr peft.mappingr rr otherrrrrrrr peft_typesrr r)rrrrrr$rr!rr3s#   8N9.46!p bhhX mr- -3J-ei-W-8!'!7:!NQ!!0$)# w" w  w  wv$(BZ^{%{%({%>V{% {%r!