+ j0}dRt^RIt^RIHt^RIt^RIHt^RIHt^RI H t ^RI H t ^RI Ht^RIHtHtHtHtHtHtHt^R IHt]P2!]4tRsR t!R R ]P<P>4t R#R lt!R#Rlt"!RR]PF4t$!RR]PF4t%!RR] 4t&]!RR]44t']!RR7]!RR]444t(]!RR7]!RR]444t)]!RR]'44t*]!R R7!R!R"]'] 44t+.R$Ot,R#)%zPyTorch RWKV model.N) dataclass)nn)initialization)GenerationMixin)GradientCheckpointingLayer)PreTrainedModel) ModelOutputauto_docstringis_bitsandbytes_availableis_kernels_availableis_ninja_availableis_torch_cuda_availablelogging RwkvConfigcr\4'g \R4h^RIHpV!R^R7sV\nR#)zFkernels is not installed, please install it with `pip install kernels`) get_kernelzkernels-community/rwkv)versionN)r ImportErrorintegrations.hub_kernelsrrwkv_cuda_kernelmax_seq_length)context_lengthrs& g/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/rwkv/modeling_rwkv.pyload_wkv_cuda_kernelr-s1  ! !bcc6!":AF&4#cHa]tRt^8to]RRl4t]RRl4tRtVtR#)RwkvLinearAttentionNc 4VP4wrxp V\P8d!\RV R\P R24hWy,\ V ^ 4,^8wd \RV RV R\ V ^ 4 R24hVP VnVPPR8wgRVPPR8wg7VPPR8wgVPPR8wd \R4h\P!VP4P44)pVP \P8Xd1VP4pVP4pVP4pVP4pVP4pVP4p\P!V\PR 7p V'gVeVfY\P !VV ^\P"VP\PR 7pVR ;;,R ,uu&MD\P$!VU u.uFqP'^4NK up ^R 7P4pVP \P(8Xd\P*p M\P,p V !WW4W4MIVP \P(8Xd\P.M\P0p V !WW4V 4VP3WW4V 4Ve7\P4!V^^R 7U u.uFqP7^4NK pp V P9VP 4V3#uup iuup i)zCannot process a batch with z+ tokens at the same time, use a maximum of z with this model.zThe product of batch size (z) and hidden size (z") needs to be a round multiple of .cudazUCalling the CUDA kernel for wkv attention requires all tensors to be on CUDA devices. memory_format)dtypedevicer"籡*G)dim)NNNr')sizerr ValueErrorminr# input_dtyper$typetorchexpfloat contiguousfloat16 empty_likecontiguous_formatzerosfloat32cat unsqueezebfloat16forward_with_state_bf16forward_with_state forward_bf16forwardsave_for_backwardchunksqueezeto) ctx time_decay time_firstkeyvaluestate return_state batch_sizeseq_len hidden_sizeoutputs forward_funcs &&&&&&& rr=RwkvLinearAttention.forward9s+.88:( [ %44 4.wi7b#2233DF   #c+r&: :a ?-j\9L[MZ";34A7  ))    " "f ,  %%/zz&(||  F*tu uii 0 0 2 = = ?@@ 99 %#))+J))+CKKME**, nn  "!!#U5L5LM 5,} --::"'"9"9 g$& 5"A5a;;q>5"AqITTVyyENN*/GG /BB V K>$A%..u}} '' %BYBYZ   E4K4KL""58O8OP %-- '~~'H:E:W(66]m]v]v          !      OOK ( OOK ( HH[ ! JJ{ #     rNFN) __name__ __module__ __qualname____firstlineno__ staticmethodr=rS__static_attributes____classdictcell__ __classdict__s@rrr8s,<1<1|% % rrcVP4wrgp\P!V4pVf\P!VR,\PR7p \P!VR,\PR7p \P!VR,\PR7R, p MVwrp \P!V4)p\ V4EFXp VRV 3,P 4p VRV 3,p\P!WV,4p\P!W, 4p\P!W,V, 4pVV ,VV,,pVV ,V,pVV, PVP4VRV 3&\P!W,V 4p\P!W,V, 4p\P!V V, 4pVV ,VV,,p VV ,V,p Tp EK[ V'gVeWV .pW3#)Nr')r#r%r') r)r. zeros_liker6r/ranger0maximumrAr#)rCrDrErFrGrH_ seq_lengthrL num_state den_state max_state current_index current_key current_valuemax_for_outpute1e2 numerator denominator max_for_states&&&&&& rrwkv_linear_attention_cpur|sxxzA1   c "F }$$SYemmD $$SYemmD $$SYemmDtK */' i ))J''Jz* !]*+113 a./ y 2JK YYy1 2 YY{/.@ ANR-%77 9nr) $- $;#?#? #Mq-  i& YY{]2 3NR-%77 NR' ! %+(u(y1 =rc \;QJdRWW#34F 'gK RM RM!RWW#344pVP^4^8Hp\eV'g V'd\WW#WER7#\P WW#WE4#)c3R"TFqPPR8gxK R#5i)r N)r$r-).0ts& r (rwkv_linear_attention..sX3Wa((--6)3Ws%'TFrGrH)anyr)rr|rapply)rCrDrErFrGrHno_cuda one_tokens&&&&&& rrwkv_linear_attentionrsocXJC3WXcccXJC3WXXG q I7i(SXtt"((UaarcJaa]tRt^toRV3RlltRRltRRltRtVtV;t #)RwkvSelfAttentionc<\SV`4Wn\RJ;'d\PVP 8Hp\ 4'd/\4'dV'g\VP 4W n VPpVPe VPMTpWPn\P !\"P$!V44Vn\P !\"P$!V44Vn\P !\"P$!^^V44Vn\P !\"P$!^^V44Vn\P !\"P$!^^V44Vn\P0!R4Vn\P4!WERR7Vn\P4!WERR7Vn\P4!WERR7Vn\P4!WTRR7VnR# \d\PR4ELi;i)Nz9Could not load the custom CUDA kernel for RWKV attention.Fbiasrjrj)super__init__configrrrr r r Exceptionloggerinfolayer_idrKattention_hidden_sizer Parameterr.emptyrCrD time_mix_keytime_mix_valuetime_mix_receptance ZeroPad2d time_shiftLinearrErF receptancerL)selfrr kernel_loadedrKr __class__s&&& rrRwkvSelfAttention.__init__s  (4qq9I9X9X\b\q\q9q   $;$=$=m Y$V%:%:;! (( ,2,H,H,TF ( (Ze &;",,u{{3H'IJ,,u{{3H'IJLLQ;)GH ll5;;q![+IJ#%<< Aq+0N#O ,,}599[eLYY{N ))KUSii 5O ) Y WX Ys,H77 IIcVP^4^8Xd#VeV^,RRVP3,pM5VPV4pVe!V^,RRVP3,VR&WP,V^VP, ,,pWP,V^VP, ,,pWP ,V^VP , ,,pVP V4pVPV4p\P!VPV44pVe!VR,V^,RRVP3&WdWR3#rr'ri)r'r) r)rrrrrrErFr.sigmoidr)rhiddenrGshiftedrErFrs&&& rextract_key_value#RwkvSelfAttention.extract_key_values* ;;q>Q 5#4Ahq!T]]23Goof-G %aAt}})< = (((7a$:K:K6K+LL,,,w!d>Q>Q:Q/RR666AH`H`D`9aa hhsm 5!]]4??:#>?  ,25ME!HQ4==( ),,rc aSPWR7wrErbVeA\;QJd.V3RlVR,4FNK 5M!V3RlVR,44MRp\SPSPVVVVR7wrVeaV^,V^,RRSP 3&V^,V^,RRSP 3&V^,V^,RRSP 3&SP WH,4V3#)rGNc3N<"TFqRRSP3,xK R#5i)r'Nr)rrMrs& rr,RwkvSelfAttention.forward.. s"FIqaDMM122Is"%:r(NNrr')rtuplerrCrDrrL) rrrG use_cacherrErF layer_staterwkvs f&&& rr=RwkvSelfAttention.forward s(,(>(>v(>(S% JOJ[eeFE"IFeeFE"IFFae 1 OO OO  "    ",7NE!HQ4==( ),7NE!HQ4==( ),7NE!HQ4==( ){{:,-u44r) rrrErrLrrCrDrrrrrFrjr^r]) r_r`rarbrrr=rdre __classcell__rrgs@@rrrsP<-&55rrc@aa]tRtRtoRV3RlltRRltRtVtV;t#)RwkvFeedForwardicV<\SV`4WnW nVPpVP e VP M^VP,p\ P!R4Vn\ P!\P!^^V44Vn \ P!\P!^^V44Vn \ P!W4RR7Vn\ P!W3RR7Vn\ P!WCRR7VnR#)NFrr)rrrrrKintermediate_sizerrrrr.rrrrrErrF)rrrrKrrs&&& rrRwkvFeedForward.__init__ s   (( (.(@(@(LF $ $RSV\VhVhRh ,,}5LLQ;)GH#%<< Aq+0N#O 99[%H))K5IYY0EJ rcVP^4^8Xd#VeV^,RRVP3,pM5VPV4pVe!V^,RRVP3,VR&WP,V^VP, ,,pWP,V^VP, ,,p\ P !\ P!VPV444pVPV4p\ P!VPV44pVe!VR,V^,RRVP3&WV,V3#r) r)rrrrr.squarerelurErFrr)rrrGrrErrFs&&& rr=RwkvFeedForward.forward1s ;;q>Q 5#4Ahq!T]]23Goof-G %aAt}})< = (((7a$:K:K6K+LL666AH`H`D`9aa ll5::dhhsm45 3]]4??:#>?  ,25ME!HQ4==( )!5((r)rrErrrrrrFrr^ r_r`rarbrr=rdrerrs@@rrrsK"))rrc<aa]tRtRtoV3RltRRltRtVtV;t#) RwkvBlockiEc<\SV`4WnW nV^8Xd2\P !VP VPR7Vn\P !VP VPR7Vn \P !VP VPR7Vn \W4Vn \W4VnR#)rj)epsN)rrrrr LayerNormrKlayer_norm_epsilonpre_lnln1ln2r attentionr feed_forward)rrrrs&&&rrRwkvBlock.__init__Fs   q=,,v'9'9v?X?XYDK<< 2 28Q8QR<< 2 28Q8QR*6<+F=rc8VP^8XdVPV4pVPVPV4W#R7wrRW,pVP VP V4VR7wrbW,pW3pV'd Wu3, pV#VR, pV#)rj)rGrrr^)rrrrrr)rrrGroutput_attentionsrroutputss&&&&& rr=RwkvBlock.forwardTs ==A [[(F>>$((6*:%>] #"//0@/N &/  | #G w Gr)rrrrrrr)NFFrrs@@rrrEs >rrcaa]tRtRtoRtR.tRR.tRtRt] P!4V3RlV3Rll4t V3R lt R t VtV;t#) RwkvPreTrainedModeligrrrCrDTc4<V^8dQhRS[P/#)r(module)rModule)formatrgs"r __annotate__ RwkvPreTrainedModel.__annotate__qsG7G7BIIG7rc  : <\SV`V4\V\4'EdVPpVP P pVP PpVPpW#^, , pRW#, , p\P!\V4Uu.uF qV, NK upVPPVPPR7p V R ,p \V4U u.uF5p R ^W^, , RRV,,,,,NK7 p p \P!WPPVPPR7p \P!\V4Uu.uFq^,^,^, NK upVP PVP PR7R,p \"P$!VPV 4\"P$!VP \P&!VP \(P*!R4,V ,44\"P$!VP\P,!W44\"P$!VP.\P,!W4RV,,4\"P$!VP0\P,!V RV,44R#\V\24'EdVPpVP P pVP PpRW#, , p\P!\V4Uu.uF qV, NK upVPPVPPR7p V R ,p \"P$!VP\P,!W44\"P$!VP0\P,!W44R#\V\4P64'dVP8P:p RpRpVP<e!\"P>!VP<4V ^,V ^,8d,\(P@!V ^,V ^,, 4pV ^,VP PB8Xd%V ^,VP P8XdRpW,p\"PD!VP8VR7R#\V\4PF4'dpVP8P:p R \(P@!\IV ^,V ^,44,p\"PD!VP8VR7R#R#uupiuup iuupiuupi) zInitialize the weights.g?)r#r$Ngffffff?g?g?g333333?)gaing-C6?)NNr')%r _init_weights isinstancerrrnum_hidden_layersrKrr.tensorrlrr#r$rCrDinitcopy_ ones_likemathlogpowrrrrrweightshaperzeros_sqrt vocab_size orthogonal_ Embeddingmax)rrrrrKr ratio_0_to_1ratio_1_to_almost0i time_weighth decay_speedzigzagrrscalers&& rr!RwkvPreTrainedModel._init_weightspsO f% f/ 0 0H & ? ?  --33K$*$@$@ !#1'<=L!$(D!E ,,*/ *<=* L*>s'X ;XX&Xc&<V^8dQh/S[;R&#)r(rr)rrgs"rrrgs rr\)r_r`rarbbase_model_prefix_no_split_modules_keep_in_fp32_modulessupports_gradient_checkpointing _is_statefulr.no_gradr__annotate_func__rdrerrs@@rrrgsR$ )<8&*#L ]]_G7G7G7rrz+ Class for the RWKV model outputs. ) custom_introc@a]tRtRtoRtRtRtRtRtV3Rlt Rt Vt R#) RwkvOutputiz state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`): The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to avoid providing the old `input_ids`. Nc <V^8dQh/S[PR,;R&S[S[P,R,;R&S[S[PR3,R,;R&S[S[PR3,R,;R&#)r(Nlast_hidden_staterG. hidden_states attentionsr. FloatTensorlistr)rrgs"rrRwkvOutput.__annotate__s~((4/6 !! "T )0**C/047> e'',-4;!rr\) r_r`rarb__doc__rrGrrrrdrerfs@rrrs( 37,0E:>M7;J!rrzK Base class for causal language model (or autoregressive) outputs. cDa]tRtRtoRtRtRtRtRtRt V3Rlt Rt Vt R#)RwkvCausalLMOutputiaP loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction). logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`): The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to avoid providing the old `input_ids`. Nc8<V^8dQh/S[PR,;R&S[PR,;R&S[S[P,R,;R&S[S[PR3,R,;R&S[S[PR3,R,;R&#)r(NlosslogitsrG.rrr )rrgs"rrRwkvCausalLMOutput.__annotate__s"   d ")#$    $+%& !! "T )0'(**C/047>)*e'',-4;+rr\) r_r`rarbr rrrGrrrrdrerfs@rrrs,&*D'+F,0E:>M7;J+rrcjaa]tRtRtoV3RltRtRt]R V3RlRll4tRt Rt R t Vt V;t #) RwkvModelic <\SV`V4\P!VPVP 4Vn\P!\VP4Uu.uFp\WR7NK up4Vn \P!VP 4Vn RVnRVnVP!4R#uupi)rFN)rrrrrrK embeddings ModuleListrlrrblocksrln_outlayers_are_rescaledgradient_checkpointing post_init)rridxrs&& rrRwkvModel.__init__s  ,,v'8'8&:L:LMmmPUV\VnVnPo$pPoYv%DPo$pq ll6#5#56 #( &+# %qs(CcVP#r^rrs&rget_input_embeddingsRwkvModel.get_input_embeddingss rcWnR#r^r!rnew_embeddingss&&rset_input_embeddingsRwkvModel.set_input_embeddingss(rc0<V^8dQhRS[PR,RS[PR,RS[PR,RS[S[P,R,RS[R,RS[R,RS[R,R S[R,R S[S[,/ #) r( input_idsNattention_mask inputs_embedsrGrroutput_hidden_states return_dictreturn)r. LongTensorr r boolrr)rrgs"rrRwkvModel.__annotate__sh h ##d*h ((4/h ((4/ h E%%&- h $; h  $;h #Tkh D[h   h rc nVeTMVPPpVeTMVPPpVeTM)VP'gVPPMRpVeTMVPP pVe\ PR4VPVP8XdVP4VeVe \R4hVfVf \R4hVfVPV4pV'dVfVP^4VPPVPP3p \^4U u.uFFp \ P"!T RV ^8:d VP$M\ P&RVP(/NKH pp V^;;,R,uu&VP*'d2VP'd V'd\ PR 4RpTp V'dR MRp V'dR MRp\-VP.4FwppV!WWVR 7wrpVP'dNVPP0^8d3V^,VPP0,^8Xd V ^, p V'd W3,pV'gKV V3,p K VP3V 4p V'd W3,pV'g1\4;QJd.R WW34FNK 5#!R WW344#\7V VVV R 7#uup i)a( input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()` (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary. If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) state (tuple of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`, *optional*): If passed along, the model uses the previous state in all the blocks (which will give the output for the `input_ids` provided as if the model add `state_input_ids + input_ids` as context). use_cache (`bool`, *optional*): If set to `True`, the last state is returned and can be used to quickly generate the next logits. NFz<`attention_mask` was passed, but it is unused in this model.zDYou cannot specify both input_ids and inputs_embeds at the same timez5You have to specify either input_ids or inputs_embedsr#r$gꌠ9Y>)FzZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...)rGrrc30"TF qfKVxK R#5ir^r\)rxs& rr$RwkvModel.forward..^st$bq$bs )rrGrrr\)rrr.trainingrr/r warning_oncer_rescale_layersr*rr)rKrrlr.r5r#r6r$r enumerater rescale_everyrrr)rr+r,r-rGrrr.r/kwargsrrrall_self_attentionsall_hidden_statesrblockrs&&&&&&&&&, rr=RwkvModel.forwards@2C1N-TXT_T_TqTq$8$D $++JjJj "+!6IZ^ZgZgZgT[[=R=Rmr %0%cd d  =#8TU U   OOI6M "''*DKK,C,CT[[EbEbcE q "A 9:a-"5"5U]][h[o[o"  !H H  & & &4===##p" % $5b4"6BD#DKK0JC/4i0 ,M* (((KK--11W 9 99Q> - 1 #$58H$H!  &9ZM&I#!1$ M2  14D D 5t];L$bt5 t5t];L$btt t++*   [sA L2c @VPVP'*8XdR#VPP^8Ed7\P !4;_uu_4\ VP4EFwrVP'dVPPPP^\WPP,4,4VPPPP^\WPP,4,4K\VPPPR4'dVPPPP P#^\WPP,4,4VPPPP P#^\WPP,4,4EK\VPPPR4'dPVP%VPPV4VP%VPPV4EKDVPPPP#^\WPP,4,4VPPPP#^\WPP,4,4EK RRR4VP'*VnR# +'giL(;i)NSCB quant_state)rr8rr<r.rr;rrrLrmul_intrrFhasattrrCdiv_ _bnb_4bit_dequantize_and_rescale)rblock_idr@s& rr:RwkvModel._rescale_layersgsC  # #DMM(9 :  ;; $ $q ('0'=OH}}}..55::1HP[P[PiPiDi@j;jk**0077<>qCT_T_TmTmHmDn?no!..44;;@@c(VaVaVoVoJoFpApq(>!"(,}}#4 #!s LN  N c \4'g \R4h^RIpVPP VP P VP P4pVP^\W PP,4,4VPPVPR4RR7PVP4p\!VRV4R#)z Perform the dequantization and rescaling of the weights of a given layer. After that operation the layer will be quantized again. z/Please install bitsandbytes to use this method.NcpuF) requires_gradr)r r bitsandbytes functionaldequantize_4bitrdatarDrHrFrr<r Params4bitrAr$setattr)r target_layerrJbnbdequant_weights quant_weights&&& rrI*RwkvModel._bnb_4bit_dequantize_and_rescales )**OP P"..889L9L9Q9QS_SfSfSrSrsQ#h++2K2K&K"LLMvv((););E)BRW(X[[\k\r\rs  h 5r)rrrrr)NNNNNNNN)r_r`rarbrr#r(r r=r:rIrdrerrs@@rrrs@ )h h h T5066rrz The RWKV Model transformer with a language modeling head on top (linear layer with weights tied to the input embeddings). cfaa]tRtRtoRR/tV3RltRtRt]R V3RlRll4t R t Vt V;t #) RwkvForCausalLMiz head.weightzrwkv.embeddings.weightc<\SV`V4\V4Vn\P !VP VPRR7VnVP4R#)FrN) rrrrrrrKrheadr)rrrs&&rrRwkvForCausalLM.__init__sH  f% IIf00&2C2C%P  rcVP#r^r]r"s&rget_output_embeddings%RwkvForCausalLM.get_output_embeddingss yyrcWnR#r^r`r&s&&rset_output_embeddings%RwkvForCausalLM.set_output_embeddingss" rc<V^8dQhRS[PR,RS[PR,RS[PR,RS[S[P,R,RS[PR,RS[R,RS[R,R S[R,R S[R,R S[S[P ,R S[S[,/ #) r(r+Nr,r-rGlabelsrrr.r/logits_to_keepr0) r.r1r r r2rFTensorrr)rrgs"rrRwkvForCausalLM.__annotate__sD D ##d*D ((4/D ((4/ D E%%&- D   4' D $;D  $;D #TkD D[D ell*D  # #D rc V eT MVPPp VPVVVVVVV R7p V ^,p \V \4'd\ V )R4MT pVP V RVR3,4pRpVe-VP!R RVRVRVPP/V BpV 'g"V3V R,,pVe V3V,#T#\VVV PV PV PR7#) a input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values.get_seq_length()` (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary. If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) state (tuple of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`, *optional*): If passed along, the model uses the previous state in all the blocks (which will give the output for the `input_ids` provided as if the model add `state_input_ids + input_ids` as context). labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` use_cache (`bool`, *optional*): If set to `True`, the last state is returned and can be used to quickly generate the next logits. N)r-rGrrr.r/r'rrgr:rNN)rrrGrrr\) rr/rrrFslicer] loss_functionrrrGrr)rr+r,r-rGrgrrr.r/rhr= rwkv_outputsr slice_indicesrrrLs&&&&&&&&&&&, rr=RwkvForCausalLM.forwardsL&1%SV8W8W~ot4]k =M1)<=>  %%pVpFpt{{OeOepiopDYb!11F)-)9TGf$ Ev E!$$&44#..   r)r]r) NNNNNNNNNrj) r_r`rarb_tied_weights_keysrrardr r=rdrerrs@@rr[r[sD()AB#D D D D rr[r])r[rr)-r r dataclassesrr.rrr generationrmodeling_layersrmodeling_utilsrutilsrr r r r r rconfiguration_rwkvr get_loggerr_rrrautogradFunctionrr|rrrrrrrrrr[__all__r\rrr}s} ! &)9-+   H %5g %..11g T)XbC5 C5L#)bii#)L*DP7/P7P7f   < <  <  <<  <$k6#k6k6\ V )?V  V r Br