sgU |dZddlZddlmZddlmZmZddlZddlm Z ddl Z ddl m Z ddlmZmZddlZddlmZmZdd lmZerddlZer dd lmZdd lmZej:eZ dd Z d ee!de jDdee!e jDfde!dee!e jDff dZ#dZ$dZ%dZ&dZ'y)z!PyTorch - Flax general utilities.N)UnpicklingError)DictTuple) from_bytes) flatten_dictunflatten_dict)is_safetensors_availableis_torch_available)logging) safe_open) load_filec~|stjj|}tj d||j drCi}t |d5}|jD]}|j|||< dddn` ddl }ddl m } | rd d ini} |j|fd d i| }tj dt!d|j#Dddt%||} | St'||} | S#1swY%xYw#ttf$rtjd wxYw)z(Load pytorch checkpoints in a flax modelzLoading PyTorch weights from .safetensorsflax) frameworkNrr #is_torch_greater_or_equal_than_1_13zLoading a PyTorch model in Flax, requires both PyTorch and Flax to be installed. Please see https://pytorch.org/ and https://flax.readthedocs.io/en/latest/installation.html for installation instructions. weights_onlyT map_locationcpuzPyTorch checkpoint contains c3<K|]}|jywN)numel).0ts [/var/www/html/venv/lib/python3.12/site-packages/transformers/modeling_flax_pytorch_utils.py z=load_pytorch_checkpoint_in_flax_state_dict..Ns:e1779:es,z parameters.)ospathabspathloggerinfoendswithr keys get_tensortorch pytorch_utilsr ImportErrorModuleNotFoundErrorerrorloadsumvalues"convert_pytorch_state_dict_to_flax*convert_pytorch_sharded_state_dict_to_flax) flax_modelpytorch_checkpoint_path is_shardedallow_missing_keyspt_path pt_state_dictfkr(rweights_only_kwargflax_state_dicts r*load_pytorch_checkpoint_in_flax_state_dictr<1sM ''//"9: 3G9=>   N +M7f5 77A'(||AM!$7 7 7 N<_.$!7df &EJJwYUYFXYM KK6s:emNbNbNd:e7efg6hhtu v<]JW EE\^hi 1 7 7 !45  %   s*D  D D&D< pt_tuple_key pt_tensorrandom_flax_state_dict model_prefixreturncdttdtffd }|dddz}|ddvr ||r||fS|dddz}|dd k(r ||s||fS|ddd z}|dd k(r ||s||fS|ddd z}|dd k(r ||r||fS|dddz}|dd k(r/|jdk(r ||s|j dddd}||fS|dddz}|dd k(r||s|j }||fS|dddz}|ddk(r||fS|dddz}|ddk(r||fSd}|ddddk(r |ddz}n|ddddk(r|ddz}| |dd|fz}||fS||fS)zYRename PT weight names to corresponding Flax weight names and reshape tensor if necessarykeyrAcDtt|f|zhzdkDS)zAChecks if `key` of `(prefix,) + key` is in random_flax_state_dictr)lenset)rCr@r?s ris_key_or_prefix_key_in_dictzCrename_key_and_reshape_tensor..is_key_or_prefix_key_in_dict_s+3-.# 7L1MMNQRRRN)scale)weightgamma)mean running_mean)var running_var) embeddingrK)kernelr rrKrL)biasbeta)parametrizations original0_g)rZ original1_v)rstrboolndim transposeT)r=r>r?r@rGrenamed_pt_tuple_keynames `` rrename_key_and_reshape_tensorrgWsXS%*SS (,z9B..3OPd3e#Y..(,y8B>)2N|2\#Y..(,x7B=(1Ml1[#Y..(,~=B8#(DEY(Z#Y..(,{:B8# !(;D`amDn''1a3 #Y..(,{:B8#,H,VKK #Y..(,{:B7"#Y..(,y8B6!#Y.. DBEE??B$& be!e  A AB$& +CR0D7:#Y..  ""rHc txr:ttt|j t j }|rt jnd}|jDcic]\}}||j}}}|rX|jD]E\}}|j|k(r|j}|jj||<G|j}d|jvr|jd}n |j}t|} d|jvr)t|jd} | j!| i} ||vxr2||j#Dchc]}|j%ddc}v} ||vxr2||j#Dchc]}|j%ddc}v} |jD]}\}}t'|j%d}|||k(}|d|k(}| r|r|dd}t)||| |\}}|f|z| v}| r|r|f|z}|| vrH|j*| |j*k7r,t-d|d| |j*d |j*dd|jvrd |d vsd |d vrt/j0|| d |z<d|d vr| j3|d|st/j0|n$t/j0|t.j| d|z<?|st/j0|n$t/j0|t.j| |<t5| Scc}}wcc}wcc}w)Nbfloat16params batch_stats.rr 1PyTorch checkpoint seems to be incorrect. Weight  was expected to be of shape , but is rMrIrOrknum_batches_trackeddtyperj)r isinstancenextiterr/r(Tensorriitemsrsfloatrnumpybase_model_prefixrjrupdater&splittuplergshape ValueErrorjnpasarraypopr)r7r2from_binrir9v weight_dtypesr@flax_model_paramsr?flax_batch_statsr;$load_model_with_head_into_base_model$load_base_model_into_model_with_headpt_keyr>r= is_bfloat_16has_base_model_prefixflax_key flax_tensorrequire_base_model_prefixs rr0r0s!#d 4]=Q=Q=S8T3UW\WcWc(dH!)u~~zH,9,?,?,ABDAqQZBMB!'') /DAqww("GGI uuw}}M!   / //L:$$$&--h7&--)*;< )))' (9(9-(HI%%&67O,8@Q,Q,-2D2D2FGQaGG)-9 )%;\! + &2Oh$>BX$X! /4M$1H - -  $:8$D$J$JJ GxOl-h7==>i HYHYGZZ[] J-- -"%(2,)>?B{{;?W 08 ;<$ 4##Hd31= K(#++kadamamBn K(2 3 1= K(#++kadamamBn H %U,\ / **_C4HLs.M(5M.-M3c ddl}ddlm}i}|D]O}|rddini}|j|fi|}|j D cic]\}} || j } }} |j D cic]M\}} || j |j k7r| jn| jjO}}} |j} d|jvrB|jd} t| } | jt|jdn|j} t| } | | vxr2| |jDchc]}|jddc}v}| | vxr2| |jDchc]}|jddc}v}|j D]\}}t|jd}| ||j k(}|d| k(}|r|r|dd}t!||| | \}}| f|z| v}|r|r| f|z}|| vrH|j"| |j"k7r,t%d |d | |j"d |j"dd|jvrd |d vrt'j(||d|z<d|d vrt'j(||d|z<d|d vr|j+|d#|st'j(|n$t'j(|t&j |d|z<g|st'j(|n$t'j(|t&j ||<Rt-|Scc} }wcc} }wcc}wcc}w)Nrr rrTrkrjrlrmrnrorMrIrprOrqrrrt)r(r)rr-ryrsrir{rzr|rjrr}r&r~rrgrrrrrr)shard_filenamesr2r(rr; shard_filer:r7r9rrr@rr?rrrr>r=rrrrrs rr1r1sBO%K 7Znd3`b" :D1CD 0=0C0C0EF1AGGF FYfYlYlYn QUQRTUAAGGu~~5qwwy1779??;L L  "33  J-- - * 1 1( ; %12C%D " " ) ),z7H7H7W*X Y * 1 1 %12C%D "0x(H(N(NN$KF8Sp1(;AAB)KL]L]K^^_a  1 11Xb\)CF;;{C[O$4x$?@HRL(CF;;{C[O$4x$?@(HRL8#''$75ACKK ,ckkR]eheqeqFr h 675ACKK ,ckkR]eheqeqFr )[/ 9KX / **QG "L!PsM-.AM3M9M>ctjj|}tj d|t t d|jjz}|jdrt|}t|d}n1t|d5} t||j} d d d t#|S#t$rt!d|dwxYw#1swY1xYw) (Load flax checkpoints in a PyTorch modelzLoading Flax weights from Flaxrrl)seprbzUnable to convert z to Flax deserializable object. N)r r!r"r#r$getattr transformers __class____name__r%safe_load_fileropenrreadrEnvironmentError"load_flax_weights_in_pytorch_model)modelflax_checkpoint_pathflax_clsr;state_fs r%load_flax_checkpoint_in_pytorch_modelrQs77??+?@ KK,-A,BCD|Veoo.F.F%FGH$$^4()=>(cB & - t t",Xw||~"F t .e_ EE# t&);z4load_flax_weights_in_pytorch_model..usCLLAXrHzFound ``bfloat16`` weights in Flax model. Casting all ``bfloat16`` weights to ``float32`` before loading those in PyTorch model.c~|jtjk(r|jtj S|Sr)rsrriastypenpfloat32rts rrz4load_flax_weights_in_pytorch_model..~s'  8T6==4Z`rHrlr rIrRrSrV)rUrTrr )rJrQrM)rNrO)rPrkrYrTrZr[r\r]r^r_z.Flax checkpoint seems to be incorrect. Weight rnrozQSome weights of the Flax model were not used when initializing the PyTorch model z: z, - This IS expected if you are initializing z from a Flax model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a FlaxBertForPreTraining model). - This IS NOT expected if you are initializing z from a Flax model that you expect to be exactly identical (e.g. initializing a BertForSequenceClassification model from a FlaxBertForSequenceClassification model).z3All Flax model weights were used when initializing z. zSome weights of zE were not initialized from the Flax model and are newly initialized: zo You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.zAll the weights of z were initialized from the Flax model. If your task is similar to the task the model of the checkpoint was trained on, you can already use z* for predictions without further training.)%r(r*r+r#r,rjax tree_utiltree_mapr/anywarning state_dictr|r&r~rFryjoinrbrrcrdrrrurndarrayr from_numpyremoveappendload_state_dictlistrErr)pt_model flax_stater( is_type_bf16r; pt_model_dictr9rrunexpected_keys missing_keysflax_key_tuplerrrrspecial_pt_namesrCkey_componentsrf key_to_checks rrrgs  6 67XZd efmmoL <  5 ]]++ `bl #:.O'')M,4,F,F*,T,""MDVDVDX*Yq1773<?*YY)-5,F,Fj,X,"" @R@R@T&U1qwws|A&UU) O}))+,L'6'<'<'>A-# .q 1X5O5O O$'HHh.H.H-J^-[$\`m$m! 04I+AB/N 16O&88:^KN "  )k.>.>!.CQ_H`huHu+CR0;>N-- \BK B 8 +0HP]0]+CR0;>N%--K B #9 9+CR0;>N~b) )+CR03DDN nR( (+CR03CCN J &xxqr 23Hxx/H 5C YYs^NDbe!e$);[(II%b)D0A&+={*KK%b)D0!/!4v!="xx7 14 . 5 ' ''1H } $  M($;$A$AA D^DTU&&3H&=&C&C%DIkN_N_M``ac >H UWU_U_=`bjj5fq *:%*:*:;*G h'##H-  " "8 ,CA-F ]+ %L ?a ""++,B.?@""++,-88@8J8J8S8S7TU9  9  LXM_M_MhMhLiilmn <1x11::;<)N+5 5  O !("4"4"="=!>?##+#5#5#>#>"??i k OI , -    2+Z'VsS.TT.&T)F)(__doc__r picklertypingrrr jax.numpyr{rrflax.serializationrflax.traverse_utilrrrr r utilsr r( safetensorsr safetensors.flaxrr get_loggerrr#r<r`rrgr0r1rrrHrrs( " );:%<   H %IN#L@#*@#zz@#!ckk!12@# @#  Cj"** @#FT+xS+vF,IrH