+ jRaR=ta0t^RIt^RIt^RIt^RIt^RIt^RIt^RIt^RIt^RI t ^RI t ^RI t ^RI H t Ht^RIHtHtHtHtHt^RIHt^RIHt^RIHt^RIHt^RIHt^R IHtHtH t H!t!^R I"H#t#^RI$t$^R I$H%t%^R I&H't'H(t(^R I)H*t*^RI+H,t,H-t-^RI.H/t/H0t0H1t1H2t2H3t3^RI4H5t5H6t6H7t7^RI8H9t9H:t:H;t;Ht>H?t?H@t@^RIAHBtB^RICHDtDHEtEHFtF^RIGHHtH^RIIHJtJ.R>OtKRtL]^kRtM]^k]!RR44tN]!RR 44tO!R!R]4tPR"tQR#R$ltR!R%R&] 4tS!R'R(]S4tT!R)R*]S4tU!R+R,4tVR-R.ltWR/R0ltXR1R2ltYR3R4ltZ!R5R] 4t[!R6R][4t\!R7R8]E4t]!R9R:4t^!R;R]D4t_!R<R]]]B4t`R#)?N)ABCabstractmethod)Callable GeneratorIterableIteratorSequence)contextmanager) dataclass)Enum)UnsupportedOperation)Path)AnycastFinalIO)Buffer)Tensor)_get_available_device_type_get_device_module)narrow_tensor_by_index)ExtensionRegistryStreamTransformExtension)CUSTOM_METADATA_KEYDCP_VERSION_KEY FORMAT_KEY FORMAT_VALUEHF_DCP_VERSION)MetadataSTATE_DICT_TYPE StorageMeta) LoadItemTypeLoadPlan LoadPlannerReadItemSavePlan SavePlanner WriteItem WriteItemType)BlockingAsyncStager) StorageReader StorageWriter WriteResult)_create_file_view)FutureFileSystemWriterFileSystemReader FileSystemFileSystemBaseSerializationFormatz .metadataz1.0.0c:a]tRt^JtoRtRtRtV3RltRtVt R#) _StorageInfoz#This is the per entry storage info.NcvVPP4UUu/uF wrVfK WbK upp#uuppiN)__dict__items)selfkvs& h/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/distributed/checkpoint/filesystem.py __getstate___StorageInfo.__getstate__Ss1!%!4!4!6H!6!!6HHHs55ch<V^8dQh/S[;R&S[;R&S[;R&S[S[,R,;R&#) relative_pathoffsetlengthNtransform_descriptors)strintr )format __classdict__s"r> __annotate___StorageInfo.__annotate__Js@  K  K $C=4/6) __name__ __module__ __qualname____firstlineno____doc__rFr?__annotate_func____static_attributes____classdictcell__rJs@r>r6r6Js- 37IrMr6c,a]tRt^WtoV3RltRtVtR#)_StoragePrefixc&<V^8dQh/S[;R&#)rBprefixrG)rIrJs"r>rK_StoragePrefix.__annotate__Ws KrMrNN)rOrPrQrRrTrUrVrWs@r>rYrYWsrMrYc]tRt^\tRtRtRtR#)r4 torch_save safetensorsrNN)rOrPrQrR TORCH_SAVE SAFETENSORSrUrNrMr>r4r4\s JKrMz.distcpc$V^8dQhR\/#rBreturnr\)rIs"r>rKrKdsrMc>\\P!44#r8)rGuuiduuid4rNrMr>_generate_uuidrids tzz| rMcta]tRt^hto]V3RlRl4t]V3RlRl4t]V3RlRl4tRtVt R#) _TensorLoaderc*<V^8dQhRS[RS[RR/#rBsizeobjreNrHobject)rIrJs"r>rK_TensorLoader.__annotate__js"   & T rMcR#r8rNr;rnros&&&r>add_TensorLoader.addi rMc<V^8dQhRR/#rBreNrN)rIrJs"r>rKrrn  t rMcR#r8rNr;s&r> start_loading_TensorLoader.start_loadingmrwrMcZ<V^8dQhRS[S[S[PS[3,,/#rdrtupletorchrrq)rIrJs"r>rKrrrs'  u||V';!<= rMcR#r8rNr|s&r>values_TensorLoader.valuesqrwrMrNN) rOrPrQrRrrur}rrUrVrWs@r>rkrkhs?      rMrkcha]tRt^vtoV3RlRltV3RlRltV3RlRltV3RlRltR tVt R #) _SerialCpuLoaderc$<V^8dQhRS[RR/#)rB resolve_funreN)r)rIrJs"r>rK_SerialCpuLoader.__annotate__ws22H22rMc Wn.VnR#r8)rr:)r;rs&&r>__init___SerialCpuLoader.__init__ws&/1 rMc*<V^8dQhRS[RS[RR/#rmrp)rIrJs"r>rKr{s"'''&'T'rMc>VPPW34R#r8)r:appendrts&&&r>ru_SerialCpuLoader.add{s 4+&rMc<V^8dQhRR/#ryrN)rIrJs"r>rKr~rzrMcR#r8rNr|s&r>r}_SerialCpuLoader.start_loading~s rMcZ<V^8dQhRS[S[S[PS[3,,/#rdr)rIrJs"r>rKrs'  u||V';!<= rMc#"VPFwwrVPV4P4pVP4pVP 4P 4VP 8wdVP4pVV3xKy R#5ir8)r:rdetachcpuuntyped_storagernnbytescloner;_rotensors& r>r_SerialCpuLoader.valuessrjjFA%%c*113FZZ\F%%',,.&--?  !sB B )r:rN) rOrPrQrRrrur}rrUrVrWs@r>rrvs-22''    rMrca]tRt^toRV3RlRllt]V3RlRl4tV3RlRltV3RlR ltV3R lR lt V3R lR lt V3RlRlt V3RlRlt Rt VtR#)_OverlappingCpuLoaderNcR<V^8dQhRS[RS[PR,RS[RR/#)rBrstreamNinflight_threshholdre)rrStreamrH)rIrJs"r>rK"_OverlappingCpuLoader.__annotate__sBIII t#I! I  IrMcVWn.VnW0n^Vn\P !4Vn^VnRVnV'd VPM \4Vn \VP4Vn \\PP T;'gVPP#44VnVP$VPP#48wd6VP$P'VPP#44R#R#)rFN)rr:rin_flight_data collectionsdeque current_itemsidxstarted device_typerr device_modulerrcudarcurrent_streamr wait_stream)r;rrrs&&&&r>r_OverlappingCpuLoader.__init__s '/1 #6 0;0A0A0C "(F  .H.J 00@0@A JJ  vLL););)J)J)L   ;;$,,;;= = KK # #D$6$6$E$E$G H >rMc <V^8dQhRS[/#rdbool)rIrJs"r>rKrs++t+rMcFVP\VP48#r8)rlenr:r|s&r>_done_OverlappingCpuLoader._donesxx3tzz?**rMcZ<V^8dQhRS[S[S[PS[3,,/#rd)listrrrrq)rIrJs"r>rKrs'U5<<#789rMc.pVPVP8dVPP4VPVP8dxVPP 4pV;PV^,P 4V^,P4,,unVPV4KV#)r) rrr synchronizerpopleftnumel element_sizer)r;drainedvals& r>_drain_OverlappingCpuLoader._drains   $":": : KK # # %!!T%=%==$$,,.C   3q6<<>CF4G4G4I#I I  NN3 rMc<V^8dQhRR/#ryrN)rIrJs"r>rKrsNNNrMcVPPVP4;_uu_4VP'EgVPVP8EddVP VP ,wrV;P ^, unVPV4P4pVPPVP8XdVPRRR7pMwVP\P!R48XdSVP4P4VP!4VP",8wdVP%4pVP&P)VV34V;PVP!4VP+4,, unEKRRR4R# +'giR#;i)rT)device non_blockingN)rrrrrr:rrrrtypertorrrnritemsizerrrrrs& r>_refill_OverlappingCpuLoader._refillsC    & &t{{ 3 3jjjT%8%84;S;S%SDHH-A ))#.557==%%)9)99#YYe$YGF]]ell5&99..0557!<<>FOO;<"("")) ##v||~8K8K8M'MM##+4 3 3 3sG E?G  G cZ<V^8dQhRS[S[S[PS[3,,/#rd)rrrrrq)rIrJs"r>rKrs'""% f(<"=>"rMcVP'g \R4h\VP4^8dVPP 4VP#)z._finish called before all items were processed)rAssertionErrorrrrrr|s&r>_finish_OverlappingCpuLoader._finishsFzzz !QR R t!! "Q & KK # # %!!!rMc*<V^8dQhRS[RS[RR/#rmrp)rIrJs"r>rKrs"'''&'T'rMcxVP'd \R4hVPPW34R#)z&cannot add items after loading startedN)r RuntimeErrorr:rrts&&&r>ru_OverlappingCpuLoader.adds+ <<<GH H 4+&rMc<V^8dQhRR/#ryrN)rIrJs"r>rKrstrMcVP'dR#RVnVPP\P!^4R7VP 4R#)NTkey)rr:sortoperator itemgetterrr|s&r>r}#_OverlappingCpuLoader.start_loadings; <<<   H//23 rMcZ<V^8dQhRS[S[S[PS[3,,/#rdr)rIrJs"r>rKrs'""u||V';!<="rMc#"VP4VP'g-VP4pVP4TRjxL K>VP 4RjxL R#L L5ir8)r}rrrr)r;rs& r>r_OverlappingCpuLoader.valuessN ***kkmG LLN  <<>!! !s("A/%A/ A+ A/$A-%A/-A/) rrrrrrr:rrr)Ni@B)rOrPrQrRrpropertyrrrrrur}rrUrVrWs@r>rrscII.++NN0""'' ""rMrcLa]tRt^toRtRV3RlRlltV3RlRltRtVtR#) _StorageWriterTransforms This is experimental, and will likely move elsewhere in the future. It lives here to minimize changes while we are still learning and gathering feedback. NcB<V^8dQhRS[S[,R,RR/#)rB extensionsNre)r r)rIrJs"r>rK%_StorageWriterTransforms.__annotate__s, C C"#;r!_StorageWriterTransforms.__init__s!+ 2" rMc v<V^8dQhRS[RS[PRS[S[S[,S[S[,3,/#)rB write_item raw_streamre)r(ioIOBaserrbytesrrG)rIrJs"r>rKrs@YY#Y13Y r%y$s)# $YrMc0!RR\P4p\\\,V!V44pVP FpVP V4pK T\VP 4Uu.uFqUP4NK up3#uupi)c\a]tRtRtoV3RlRltV3RlRltV3RlRltRtR tVt R #) E_StorageWriterTransforms.transform_save_stream..NoCloseWriteric4<V^8dQhRS[P/#)rBraw)rr)rIrJs"r>rKR_StorageWriterTransforms.transform_save_stream..NoCloseWriter.__annotate__s  BII rMcWnR#r8r)r;rs&&r>rN_StorageWriterTransforms.transform_save_stream..NoCloseWriter.__init__srMc <V^8dQhRS[/#rdr)rIrJs"r>rKr s  4 rMcR#)TrNr|s&r> writeableO_StorageWriterTransforms.transform_save_stream..NoCloseWriter.writeable srMc&<V^8dQhRS[RS[/#)rBbre)rrH)rIrJs"r>rKr s ) )v )# )rMc8VPPV4#r8)rwrite)r;rs&&r>rK_StorageWriterTransforms.transform_save_stream..NoCloseWriter.write sxx~~a((rMcZVP4VPP4R#r8)flushrr|s&r>closeK_StorageWriterTransforms.transform_save_stream..NoCloseWriter.closes  rMrN) rOrPrQrRrrrrrUrVrWs@r> NoCloseWriterrs-     ) ) ! !rMr ) rrrrrr transform_toreversedget_descriptor)r;rrr r exs&&& r>transform_save_stream._StorageWriterTransforms.transform_save_streamsv !BII !BuI}Z'@A //B??<8L"Xdoo=VW=Vr002=VWXXWs7Brr8) rOrPrQrRrSrrrUrVrWs@r>rrs" C CYYrMrc0V^8dQhR\R\/#)rBitemre)r(rH)rIs"r>rKrKs 4 4Y 43 4rMc^pVPf \R4hVPPF pW,pK VPPPpV\ P PV4,#)rz&WriteItem tensor_data must not be None) tensor_datarrn propertiesdtyper_utils _element_size)rrnsrs& r> _item_sizerso D EFF    " " #    ' ' - -E %,,,,U3 33rMc~V^8dQhR\R\\,R\\\,,/#)rBbinsr:re)rHrr()rIs"r>rKrK(s.#d9o$tIBWrMcV^8XdV.#VUu.uF%q"P\P8XgK#VNK' ppVUu.uF%q"P\P8wgK#VNK' pp\V4Uu.uFp.NK pp\V4Uu.uFp^NK ppVP \ RR7\ V4F#wrWhV,,PV4K% VFhp\\ V4\P!^4R7^,p Wi,PV4Wy;;,\ V4, uu&Kj V#uupiuupiuupiuupi)rT)rreverser) rr)BYTE_IOrangerr enumeraterminrr) rr:wibytes_wtensor_wrbuckets bucket_sizesirs && r>_split_by_size_and_typer)(s% qyw!FEbWW 0E0E%ErrEGF"GUrgg1F1F&FUHG27+%>+Qb+G%>$T{+{!A{L+ MMj$M/7#D  $$)L)x/B/B1/EFqI BZ^+  N#GG%>+s" EE E%E: E$ E)cV^8dQhR\R\PR\P\P ,R\ R\R\R\/#)rB transformsrdatar storage_keyserialization_formatre) rrrBytesIOrrr(rGr4r-)rIs"r>rKrK@s_44(4 II4 **u|| #4 4  4 . 44rMc VP4pVPW14wrxVP\P8XdL\ V\ P4'g \R4hVPVP44M\ V\P4'g \R4hVP\P!R48wd \R4hV\P8Xd\P !W'4VP#4V\P8Xg!\ V\ P4'dVP4V, p M%VP%4VP'4,p \)V4^8XdRMTp \+VP,V \/VVV V R7R7#)z/Data must be io.BytesIO for BYTE_IO write itemsz5Data must be torch.Tensor for non-BYTE_IO write itemsrzTensor must be on CPU deviceN)rF)index size_in_bytes storage_data)tellrrr)r isinstancerr/rr getbufferrrrr4rasaverrrrr-r1r6) r+rr,rr-r.rDr rFrEinfo_transform_descriptorss &&&&&& r> _write_itemr9@sp[[]F,6,L,L-)\-///$ ++ !RS S4>>+,$ -- G  ;;%,,u- - !?@ @ #6#A#A A JJt *2=== bjjBB' 1 1 33 )*a/5J !   "<    rMcV^8dQhR\R\PR\PR\R\R\ R\ R\ R \R R / #) rB create_stream file_queue result_queueplannerr+r use_fsync thread_countr.reN)rqueueQueuer'rrHrr4)rIs"r>rKrKwsk k k  k ++k  k ) k  k k k .k  k rMc *VP4wrp \PP4p \ \V R4p V^8Xdb\P P 4'gV 'd6V P 4'd V^8d\VPVR7pM\VP4pV Uu.uF%qP\P8wgK#VNK' ppVFpVP\V4V4K VP4V Uu.uF%qP\P8XgK#VNK' pp.pV!V R4;_uu_4pVF3pVPV4pVP!\#VVVVV V44K5 /p/pVP%4FwppVP&'g \)R4hVP!\#VVVVV V44VVVP*P,&RVP.P0P2/VVP*P,&K V\4P68XdW^RIHpVP=V!V\>\@PB!V4\D\G\H4\J\L/R74V'd6VPO4\PPR!VPU44VP_4RRR4VPaV4EK#uupiuupi \V\X3d$p\ZP\!RT 24Rp?LeRp?ii;i +'giLe;i \bPddR#i;i) TN)rwbzTensor must be on CPU saved_offsets)r7)metadata9fsync not supported for this stream, relying on flush(): )3 get_nowaitr_C_get_privateuse1_backend_namegetattrr is_availabler resolve_datarrr)rrurr}rr9ris_cpurr1fqnrchunkoffsetsr4rbsafetensors.torchr7rrjsondumpsrrGrrrrosfsyncfilenoAttributeErrorr warningswarnrputrAEmpty)r;r<r=r>r+rr?r@r. file_namer- write_itemscustom_backend_namecustom_device_modloaderr#r%rr$ write_resultsrr, tensor_dict metadata_dictrr7es&&&&&&&&& r>_write_files_from_queuerfws` 2<2G2G2I /IK#((("H"H"J  '/BD I  !JJ++--).?.L.L.N.N'!+.(((; *((&1U[rGG}?T?T4T[HU&  :j1:>'  "$/TKb77m>S>S3SrrKGTMy$//6")J"// ;D!((#&" &'0  #*! " *0--/&FJ!===,-DEE!((#&""&'0  9?K 0 0 4 45')?)?)E)E)M)M;M*"2"2"6"67+:$(+>+J+JJ6LL' 3TZZ 5N /^1D *L& LLN1  s0t   ] +EV Up+,@A WXYWZ[k0/v ;;   sA-M90M98M98M9 L%+L%1:M9+ L*L*M9+EM&M&$L/9M& &M9/M#MM&M##M&& M6 1M99NNca]tRtRto]]V3RlRl44t]V3RlRl4t]V3RlRl4t]V3RlR l4t ]V3R lR l4t ] ]V3R lR l44t ]V3RlRl4t ]V3RlRl4tRtVtR#)r3ic z<V^8dQhRS[S[P,RS[RS[S[P RR3,/#rBpathmodereNrGrUPathLikerrr)rIrJs"r>rKFileSystemBase.__annotate__s>//"++%/-0/ 299dD( )/rMcR#r8rN)r;rjrks&&&r>r;FileSystemBase.create_streams,/rMct<V^8dQhRS[S[P,RS[RS[S[P,/#rBrjsuffixrerGrUrm)rIrJs"r>rKrns7  "++% /2 r{{  rMcR#r8rNr;rjrss&&&r> concat_pathFileSystemBase.concat_paths rMcr<V^8dQhRS[S[P,RS[S[P,RR/#rBrjnew_pathreNrt)rIrJs"r>rKrns*WW3,Wbkk8IWdWrMcR#r8rNr;rjr{s&&&r>renameFileSystemBase.renamesTWrMcn<V^8dQhRS[S[P,RS[S[P,/#rBrjrert)rIrJs"r>rKrns%JJcBKK/JC"++4EJrMcR#r8rNr;rjs&&r> init_pathFileSystemBase.init_pathsGJrMcH<V^8dQhRS[S[P,RR/#rBrjreNrt)rIrJs"r>rKrns99# +99rMcR#r8rNrs&&r>mkdirFileSystemBase.mkdirs69rMcJ<V^8dQhRS[S[P,RS[/#rB checkpoint_idrerGrUrmr)rIrJs"r>rKrnsRR33DRRrMcR#r8rNclsrs&&r>validate_checkpoint_id%FileSystemBase.validate_checkpoint_idsORrMcJ<V^8dQhRS[S[P,RS[/#rr)rIrJs"r>rKrns::3,::rMcR#r8rNrs&&r>existsFileSystemBase.existss7:rMcH<V^8dQhRS[S[P,RR/#rrt)rIrJs"r>rKrns;;C"++-;$;rMcR#r8rNrs&&r>rm_fileFileSystemBase.rm_files8;rMrNN)rOrPrQrRr rr;rwr~rr classmethodrrrrUrVrWs@r>r3r3s//  WWJJ99RR::;;rMca]tRtRto]V3RlRl4tV3RlRltV3RlRltV3RlR ltV3R lR lt ] V3R lR l4t V3RlRlt V3RlRlt V3RlRltRtVtR#)r2ic z<V^8dQhRS[S[P,RS[RS[S[P RR3,/#rirl)rIrJs"r>rKFileSystem.__annotate__s>**"++%*-0* 299dD( )*rMc#"\V\4'g \V4pVPV4;_uu_4p\\P V4xRRR4R# +'giR#;i5ir8)r5ropenrrr)r;rjrkrs&&& r>r;FileSystem.create_streamsH$%%:D YYt__ryy&) )___s;A8A$ A8$ A5 / A8ct<V^8dQhRS[S[P,RS[RS[S[P,/#rrrt)rIrJs"r>rKrs4bkk 133CTrMcT\V\4'g \V4pW, #r8r5rrvs&&&r>rwFileSystem.concat_paths $%%:D}rMcn<V^8dQhRS[S[P,RS[S[P,/#rrt)rIrJs"r>rKrs-cBKK/C"++4ErMcH\V\4'g \V4pV#r8rrs&&r>rFileSystem.init_paths$%%:D rMcr<V^8dQhRS[S[P,RS[S[P,RR/#rzrt)rIrJs"r>rKrs4**3,*bkk8I*d*rMc\V\4'g \V4pVP\\V44R#r8)r5rr~rr}s&&&r>r~FileSystem.renames+$%%:D Dx()rMcH<V^8dQhRS[S[P,RR/#rrt)rIrJs"r>rKrs$00# +00rMcn\V\4'g \V4pVPRRR7R#)T)parentsexist_okN)r5rrrs&&r>rFileSystem.mkdirs($%%:D 4$ /rMcJ<V^8dQhRS[S[P,RS[/#rr)rIrJs"r>rKr%s$  33D  rMc0\V\4'dR#R\V49dR#\V4PFRpVP 4'gK\ P !\V4\ P4'gKQR# R#)Tz://F)r5rrGrrrUaccessW_OK)rrps&& r>r!FileSystem.validate_checkpoint_id$sb mT * * C & &m$,,AxxzzbiiA88-rMcJ<V^8dQhRS[S[P,RS[/#rr)rIrJs"r>rKr2s$3,rMcd\V\4'g \V4pVP4#r8)r5rrrs&&r>rFileSystem.exists2s$$%%:D{{}rMcH<V^8dQhRS[S[P,RR/#rrt)rIrJs"r>rKr7s$C"++-$rMch\V\4'g \V4pVP4R#r8)r5runlinkrs&&r>rFileSystem.rm_file7s!$%%:D rMcZ<V^8dQhRS[S[P,RS[S[,/#r)rGrUrmr)rIrJs"r>rKr<s(00sR[[(0T#Y0rMc\V\4'g \V4pVP4Uu.uFp\V4NK up#uupir8)r5riterdirrG)r;rjrs&& r>ls FileSystem.ls<s:$%%:D $ /1A///sA rNN)rOrPrQrRr r;rwrr~rrrrrrrUrVrWs@r>r2r2sq**  ** 00     00rMc\aa]tRtRtoRtRR^RRR]P 3V3RlV3RllltR!V3RlR lltV3R lR lt V3R lR lt V3RlRlt V3RlRlt V3RlRlt V3RlRltV3RlRltV3RlRltR!V3RlRllt]V3RlRl4t]V3RlRl4tR tVtV;t#)"_FileSystemWriteriBaE Basic implementation of StorageWriter using file IO. This implementation makes the following assumptions and simplifications: * The checkpoint path is an empty or non-existing directory. * File creation is atomic The checkpoint consist of one file per write request plus a `.metadata` file with the serialized metadata. T逖Nc<V^8dQhRS[S[P,RS[RS[RS[RS[RS[RS[S[,R,R S[R S[R S[R R/ #) rBrjsingle_file_per_rank sync_filesr@per_thread_copy_ahead overwrite _extensionsNr.argskwargsre) rGrUrmrrHr rr4r)rIrJs"r>rK_FileSystemWriter.__annotate__Ps'*'*BKK'*#'* '*  '* # '*'*67$>'*2'*'*'* '*rMc  $<\S V`4\4VnVPP V4VnW nW0nW@nWPn \4Vn W`n \V4VnWnRVnRVnR#)a Initialize the writer pointing to `path`. Args: path: directory where the checkpoint will be written to. single_file_per_rank: Produce one file per rank instead of one file per tensor/blob. Default to True. sync_files : force files to be synced to permanent storage. Default to True. thread_count: Number of IO threads to use to write. Default to 1. per_thread_copy_ahead: How many bytes to copy from the GPU ahead of saving them. Default 10Mb. overwrite: Whether to allow overwriting existing checkpoints. Defaults to True. _extensions: Extensions to apply to output streams (EXPERIMENTAL) N. B. If sync_files is disabled, there's no guarantee that the checkpoint will be consistent in the case of a failure. NT)superrr2fsrrjrrr@rrisave_idrrr+r.rankuse_collectives) r;rjrrr@rrrr.rr __class__s &&&&&&&&&*,r>r_FileSystemWriter.__init__Psv6 ,GG%%d+ $8!$(%:"%' "2;?$8! $ %)rMcV<V^8dQhRS[S[P,R,RR/#rBrNrert)rIrJs"r>rKrys)((3#4t#;(t(rMctV'd!VPPV4Vn\4VnR#r8)rrrjrirr;rs&&r>reset_FileSystemWriter.resetys% ))-8DI%' rMc0<V^8dQhRS[RS[RS[RR/#)rBis_coordinatorrrreN)rr)rIrJs"r>rKr~s3CC"C+.C:=C CrMc`VPR4VnVPRR4VnR#)rrTN)getrr)r;rrrs&&*,r>set_up_storage_writer'_FileSystemWriter.set_up_storage_writer~s(JJv& %zz*;TBrMc <V^8dQhRS[/#rdr)rIrJs"r>rKrs - -$ -rMcVP'dVPRR7pMVP^R7pVPPV4#)N)r)r_get_metadata_pathrr)r; metadata_paths& r>_metadata_exists"_FileSystemWriter._metadata_existssI     333>M!333;Mww~~m,,rMc&<V^8dQhRS[RS[/#rBplanre)r&)rIrJs"r>rKrsxHrMcVPPVP4VP4'dbVP'd6\ P !RVP RVP: R2^R7M\RVP: R24hVPeBVP'g0\P!V\RVP R24R7pV#) z#Detected an existing checkpoint in z#, overwriting since self.overwrite=z. Past version 2.5 of PyTorch, `overwrite` will default to False. Set this variable to True to maintain this functionality or False to raise when an existing checkpoint is found.) stacklevelz-Checkpoint already exists and self.overwrite=.__rr3) rrrjrrrYrZrrr dataclassesreplacerYr;rs&&r>prepare_local_plan$_FileSystemWriter.prepare_local_plans  dii  " "~~~ 9$))DhY]YgYgXijkk ! #%SDNNCTTU#VWW 99 )=)=)=&&>Btyyk2C#DD rMcF<V^8dQhRS[S[,RS[S[,/#rBplansre)rr&)rIrJs"r>rKrs#hDNrMc \V4UUu.uF:wr#VPf&\P!V\ RV R24R7MTNK< pppV#uuppi)Nrrr)r!r3rrrY)r;rr(r new_planss&& r>prepare_global_plan%_FileSystemWriter.prepare_global_plansl %U+  ,  (   >Bqc)3L M ,    sAAcL<V^8dQhRS[RS[RS[S[S[,,/#)rBrr>re)r&r'r/rr-)rIrJs"r>rKrs35555 [! " 5rMca a VPo ^o V V 3Rlp\P!4pVP'di\ VP VP 4FCpV!4pVPPVPV4pVPWvV34KE MTVP FDpV!4pVPPVPV4pVPWvV.34KF VPW$4#)rcF<SP S \ 2pS^, oV#)r)r[DEFAULT_SUFFIX)r] file_count storage_plans r>gen_file._FileSystemWriter.write_data..gen_files,'../ |N;KLI !OJ rM) r3rArBrr)r@r:rrwrjr[ _write_data) r;rr>rr<bucketr]rjrrrs &&& @@r> write_data_FileSystemWriter.write_datas (,'8'8    #(++-  $ $ $1$2C2CTZZP$J ww**499i@89Q  $J ww**499i@$89# 44rMc`<V^8dQhRS[RS[PRS[S[S[,,/#)rBr>r<re)r'rArBr/rr-)rIrJs"r>rKrs7000KK0 [! " 0rMc \P!4p.p\^VP4Fp\P !\ VPPVVVVPVPVPVPVP3 R7pVP4VPV4K \ VPPVVVVPVPVPVPVPR7 VFpVP4K .pWsP!4, pK \P"d!\%4pTP'T4Tu#i;i)r)targetr) r;r<r=r>r+rr?r@r.)rArBr r@ threadingThreadrfrr;r+rrr.startrjoinrHr\r/ set_result) r;r>r<r=threadsrtresfuts &&& r>r_FileSystemWriter._write_datasB %*KKM q$++,A  .GG)) OO..OO%%--  A GGI NN1 !-$ ''//!% $ : :oo**!%!:!: A FFH ..00{{ -3XC NN3 J s7E2FFcJ<V^8dQhRS[RS[S[S[,,RR/#)rBrFresultsreN)rrr-)rIrJs"r>rKrs,$0$0x$0$tK7H2I$0d$0rMc\Vn/pVF8pTPVUu/uFqUPVPbK up4K: W1nVP 4VnVP 'g%VPeRVP \ R2M\ R2p\\VPPVPV44pVPPVR4;_uu_4p\P !W4VP"'d6VP%4\&P(!VP+44RRR4VP 'g+VPeVP5VP4p MVP54p VPP7V 4'dVPP9V 4VPP;Wz4R#uupi \,\.3d$p \0P2!RT 24Rp ? LRp ? ii;i +'giL;i)Nrz.tmprDrG)CURRENT_DCP_VERSIONversionupdater1r3 storage_metarr _metadata_fnrrrrwrjr;pickledumprrrUrVrWrXr rYrZrrrr~) r;rFr storage_mdwr_listwr tmp_filenametmp_path metadata_filerers &&& r>finish_FileSystemWriter.finishs. G   7K7Rxx87K L * $ 1 1 3'''DII,A L> . >&  dgg11$))\JK WW " "8T 2 2m KK 0##%HH]1134 3### (= 33DII>M 335M 77>>- ( ( GGOOM * x/?L"'(<=MMSTUSVW3 2s5H >9I8$HIH=8I=II I c.<V^8dQhRS[R,/#ry)r!)rIrJs"r>rKr sSSkD0SrMcD\VPVPR7#))rr)r!rrr|s&r>r_FileSystemWriter.storage_meta s););T\\RRrMcH<V^8dQhRS[R,RS[P/#rBrNrerHrUrm)rIrJs"r>rKr#'DDsTzDR[[DrMcVf\ M RV \ 2p\\VPP VP V44#Nrrrrrrwrjr;rfilenames&& r>r$_FileSystemWriter._get_metadata_path#<(, l^Btf\N:SD$''--diiBCCrMcD<V^8dQhRS[S[P,/#rdrt)rIrJs"r>rKr(sR[[0rMc VP#)zD return the checkpoint_id that will be used to save the checkpoint. rjr|s&r>r_FileSystemWriter.checkpoint_id' yyrMcJ<V^8dQhRS[S[P,RS[/#rr)rIrJs"r>rKr/(@@33D@@rMc,\PV4#r8r2rrs&&r>r(_FileSystemWriter.validate_checkpoint_id.00??rM) rrrjrrrr.rrr@r+rr8)rOrPrQrRrSr4rarrrrrrrrr'rrrrrrrUrV __classcell__rrJs@@r>rrBs  &*%/AE4G4R4R'*'*R(( CC - -(55800d$0$0LSSDD @@@rMrcLa]tRtRtoRtR V3RlRlltV3RlRltRtVtR#) _StorageReaderTransformsi3rNc2<V^8dQhRS[R,RR/#)rBextension_registryNre)r)rIrJs"r>rK%_StorageReaderTransforms.__annotate__:s!  +Vf\4VnR#TVnR#r8)rrH)r;rHs&&r>r!_StorageReaderTransforms.__init__:s!#5#=   CU rMc b<V^8dQhRS[RS[S[,RS[S[,RS[S[,/#)rB read_itemrFrre)r%r rGrr)rIrJs"r>rKrI?s?    (} uI  E rMcVPPV4pTpVF,p\V\4'gKVP V4pK. V#r8)rHfrom_descriptor_listr5rtransform_from)r;rMrFrrrPrs&&&& r>transform_load_stream._StorageReaderTransforms.transform_load_stream?sP ,,AABWX #B"677!#!2!2>!BrM)rHr8) rOrPrQrRrSrrQrUrVrWs@r>rFrF3s    rMrFcaa]tRtRtoRV3RlV3RllltV3RlRltRV3RlRlltV3RlR ltRV3R lR lltV3R lR lt V3RlRlt V3RlRlt V3RlRlt ] V3RlRl4t]V3RlRl4tRtVtV;t#)r1iMc\<V^8dQhRS[S[P,RS[R,RR/#)rBrj_extension_registryNre)rGrUrmr)rIrJs"r>rKFileSystemReader.__annotate__Ns7 $ $BKK $/5 $  $rMc<\SV`4\4VnVPP V4Vn/Vn\4Vn\V4Vn RVn RVn R#)NT) rrr2rrrjr3riload_idrFr+rr)r;rjrUrs&&&r>rFileSystemReader.__init__Ns] ,GG%%d+ ,.%' 23FG #rMc6<V^8dQhRS[RS[S[,/#)rBsinfore)r6rr)rIrJs"r>rKrV\s#TT|T5 TrMct\\\,\WPVP 44#r8)rrrr.rDrE)r;filer[s&&&r> _slice_fileFileSystemReader._slice_file\s#BuI0||U\\RSSrMcV<V^8dQhRS[S[P,R,RR/#rrt)rIrJs"r>rKrV_s)((3#4t#;(t(rMc/VnV'd!VPPV4Vn\ 4VnR#r8)r3rrrjrirXrs&&r>rFileSystemReader.reset_s- ))-8DI%' rMc:<V^8dQhRS[RS[RS[R,/#)rBrr>reN)r#r$r/)rIrJs"r>rKrVes&::h:::rMc /pVPFMpVPVP,pVPpVP V.4P V4KO VP4EFawrxVP PVPV4p VP PV R4;_uu_4p VEFp VPV P,pVPW4p VPPT VP;'gRV 4p V P\P 8XdJ\"P$!V P'R 44pVP)^4VP+W4KV P-4'dT pM6\"P$!V P'R 44pVP)^4\/\0\2P4!VRRR74p\7VV P8V P:4pVP=V 4P?4pVPA4VPA48wd;\CRV P RVPA4 RVPA4 24hVPEV4VPGV V4EK RRR4EKd \I4pVPKR4V# +'giEK;i) rbrT) map_location weights_onlyzreq z mismatch sizes z vs NrN)&r:r3 storage_indexrC setdefaultrrrwrjr;r^r+rQrFrr"rrr/readseek load_bytesseekablerrrloadrstorage_offsetslengthsresolve_tensorrrnrcopy_ commit_tensorr/r)r;rr>per_filerMitem_mdrjrCreqsr{rreq file_slicerP read_bytesrnr target_tensorrs&&& r> read_dataFileSystemReader.read_dataes|.0I$($5$5i6M6M$NG((D   b ) 0 0 ;$ $,>>#3 Mww**499mDH&&x66&C"//0A0ABG!%!1!1&!BJ%)__%J%J 55;;" &Nxx<#7#77%'ZZ0C0CB0G%H "***3;)2244'5H(*zz.2E2Eb2I'JH$MM!,!%"!JJ (-2-1""8"C$7$7")0(>(>s(C(J(J(L (--/6;;=@"0"&s'8'8&99I-J\J\J^I__cdjdododqcr s#&++F3--c=AU 76$4`h t a766sAK3BK3$D%K33 L cH<V^8dQhRS[R,RS[P/#r-r.)rIrJs"r>rKrVr/rMcVf\ M RV \ 2p\\VPP VP V44#r1r2r3s&& r>r#FileSystemReader._get_metadata_pathr6rMc,<V^8dQhRS[RS[RS[/#)rBrrre)rr)rIrJs"r>rKrVs"  3 # ( rMchVPR4pVPV4pVPPVR4;_uu_4p\P !V4pRRR4\ XRR4f\4VnVPVPn V# +'giLK;i)rreNr) rrrr;rrorKr!rrX)r;rrrrjr&rFs&*, r> read_metadataFileSystemReader.read_metadataszz&!&&t, WW " "4 . .-{{=1H/ 8^T 2 :$/MH !(, %/ .s B!! B1 c 6<V^8dQhRS[RS[RS[RS[RR/#)rBrFrrrreN)rrr)rIrJs"r>rKrVs<NN N26N?BNNQN NrMcVPVnVPR4VnVPRR4VnVPf \ R4hR#)rrTNz)storage_data must not be None in metadata)r3rrrr)r;rFrrrs&&&*,r>set_up_storage_reader&FileSystemReader.set_up_storage_readersS%11JJv& %zz*;TB    $ !LM M %rMc&<V^8dQhRS[RS[/#r)r#)rIrJs"r>rKrVsxHrMcV#r8rNrs&&r>r#FileSystemReader.prepare_local_plans rMcF<V^8dQhRS[S[,RS[S[,/#r)rr#)rIrJs"r>rKrVs#hDNrMcV#r8rN)r;rs&&r>r$FileSystemReader.prepare_global_plans rMcD<V^8dQhRS[S[P,/#rdrt)rIrJs"r>rKrVr8rMc VP#)zD return the checkpoint_id that will be used to load the checkpoint. r:r|s&r>rFileSystemReader.checkpoint_idr<rMcJ<V^8dQhRS[S[P,RS[/#rr)rIrJs"r>rKrVr>rMc,\PV4#r8r@rs&&r>r'FileSystemReader.validate_checkpoint_idrBrM)rrXrjrr3r+rr8)rOrPrQrRrr^rr|rrrrrrrrrrUrVrCrDs@@r>r1r1Ms $ $TT(( ::xDD  NN @@@rMc|aa]tRtRtoRtRR^RRRR]P 3V3RlRlltV3R lV3R lltR t Vt V;t #) r0ia Basic implementation of StorageWriter using file IO. This implementation makes the following assumptions and simplifications: * The checkpoint path is an empty or non-existing directory. * File creation is atomic The checkpoint consist of one file per write request plus a global `.metadata` file with the serialized metadata if rank coordination is enabled. a rank local `__{rank}.metadata` file with the serialized metadata if rank coordination is NOT enabled. TrFNc<V^8dQhRS[S[P,RS[RS[RS[RS[RS[RS[RS[S[,R ,R S[R R / #) rBrjrrr@rcache_staged_state_dictrrNr.re)rGrUrmrrHr rr4)rIrJs"r>rKFileSystemWriter.__annotate__s+ + BKK+ #+  +  + # + "&+ + 67$>+ 2+  + rMc  r\PVVVVVVVVV R7 \P!VVR7R#)a* Initialize the writer pointing to `path`. Args: path: directory where the checkpoint will be written to. single_file_per_rank: Produce one file per rank instead of one file per tensor/blob. Default to True. sync_files : force files to be synced to permanent storage. Default to True. thread_count: Number of IO threads to use to write. Default to 1. per_thread_copy_ahead: How many bytes to copy from the GPU ahead of saving them. Default 10Mb. cache_staged_state_dict: Whether to cache the staged state_dict. This option decreases staging latency at the cost of increased memory usage. Additionally, if this parameter is set to True, it's the expectation that the stager is maintained and reused for multiple dcp.async_save calls. Default to False. overwrite: Whether to allow overwriting existing checkpoints. Defaults to True. _extensions: Extensions to apply to output streams (EXPERIMENTAL) N. B. If sync_files is disabled, there's no guarantee that the checkpoint will be consistent in the case of a failure. )rjrrr@rrrr.)rN)rrr*) r;rjrrr@rrrrr.s &&&&&&&&&&r>rFileSystemWriter.__init__sH: "" !5!%"7#!5 #  $$ $; rMc&<V^8dQhRS[RS[/#)rB state_dictre)r )rIrJs"r>rKr s)))O)rMc 0<^Vn\SV` V4#)zOverride of AsyncStager.stage)rrstage)r;rrs&&r>rFileSystemWriter.stage s&'"w}Z((rM)r) rOrPrQrRrSr4rarrrUrVrCrDs@@r>r0r0sD "&*%/(-AE4G4R4R+ + Z)))rMc~V^8dQh/^\9d \;R&^\9d\\,;R&#)rBrr)__conditional_annotations__rGr)rIs"r>rKrKs4F cGJ*)U3Z)KrM)r0r1r2r3r4)brrrrrSrrUrrAr rgrYabcrrcollections.abcrrrrr contextlibr r enumr r pathlibrtypingrrrrtyping_extensionsrrr torch._utilsrrtorch.distributed._shard._utilsr'torch.distributed.checkpoint._extensionrr&torch.distributed.checkpoint._hf_utilsrrrrr%torch.distributed.checkpoint.metadatarr r!$torch.distributed.checkpoint.plannerr"r#r$r%r&r'r(r)$torch.distributed.checkpoint.stagingr*$torch.distributed.checkpoint.storager+r,r-"torch.distributed.checkpoint.utilsr. torch.futuresr/__all__rrr6rYr4rrirkrrrrr)r9rfr3r2rrFr1r0rK)rs@r>rs   #MM%!#''% GBYX   E A   "))  I I  I   $   C  }0X"MX"v0Y0Yf 404nk \