+ UV-jΤ^RIt^RIHtHtHt^RIHt^RIH t ^RI H t ^RI H t ^RIHtRt!RR] P"4t!R R ] P"4t!R R ] P"4t!R R] P"4t!RR] P"4t!RR] P"4t!RR] P"4t!RR] P"4t!RR] P"4t!RR] P"4tR#)N)OptionalTupleUnion)check_array_shape AudioConfig)Gemma3nRMSNormc\V4pR.V,pV^8d&\V4^8dV^,V^,3VR&V^8d&\V4^8dV^,V^,3VR&V^8d&\V4^8dV^,V^,3VR&V^8d&\V4^8dV^,V^,3VR&V#)z/Convert PyTorch padding to MLX pad_width formatrr)len)padding input_shapendim pad_widths&& m/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/mlx_vlm/models/gemma3n/audio.pyconvert_torch_to_mlx_pad_widthr s { D4I qyS\Q& WQZ0 "  qyS\Q& WQZ0 "  qyS\Q& WQZ0 "  qyS\Q& WQZ0 " ctaa]tRt^)toV3RlV3RlltV3RlRltV3RlRltV3RlRltR tVt V;t #) %Gemma3nAudioRelativePositionEmbeddingc <V^8dQhRS[/#configr)format __classdict__s"r __annotate__2Gemma3nAudioRelativePositionEmbedding.__annotate__+sII{Irc<\S V`4WnVPPVnVPP VnVP VP,VnVPP^8dVPP^, M^Vn VPPVn \P!VP VPVP,RR7VnRpRpVP ^,p\P !\#V4\#V4, 4\%V^, ^4, pV\&P(!\&P*!V4V),4,p\&P,!V4R,VnR#)rFbias?g@N)NN.)super__init__rconf_num_attention_heads num_heads hidden_sizechannelshead_dimconf_attention_context_left max_backwardconf_attention_context_right max_forwardnnLinearpos_projmathlogfloatmaxmxexparangearray_inv_timescales) selfrargskwargs min_timescale max_timescalenum_timescaleslog_timescale_incrementinv_timescales __class__s &&*, rr'.Gemma3nAudioRelativePositionEmbedding.__init__+sK  == //  7 {{66: KK 3 3a 7   ;;CC MM4>>DMM9    !+"&(( - 5#7 7# "A &#'' IIn %)@(@ @*   "xx7HrcN<V^8dQhRS[PRS[P/#)rpositionreturnr8r;)rrs"rr r!Is#++"((+bhh+rc\VP^8XgQh\P!VP\P4RR7pWP ,p\P !\P!V4\P!V4.RR7pVPV4#)raxisr ) rr8 expand_dimsastypefloat32r< concatenatesincos)r=rHdtype scaled_time timing_signals&&& r_get_timing_signal_1d_pos?Gemma3nAudioRelativePositionEmbedding._get_timing_signal_1d_posIs|}}!!!>>(//"**"=BG!5!55  VVK "&&"5 6R ##E**rcr<V^8dQhRS[PRS[RS[RS[RS[RS[RS[RS[P/#) rterm_bd_before_shift batch_sizer)num_query_blocksquery_block_sizekey_context_sizemax_span_plus_1rIr8r;int)rrs"rr r!Ssc.. hh.. .  .  ... .rcV^,V, p^V3p \P!V\WP44p V P VVVWV^,,34p V RRRRWV,13,p V P VVVVV34p V #)NNNN)r8padrshapereshape)r=rZr[r)r\r]r^r_pad_amount_last_dim padding_tupleterm_bd_paddedterm_bd_reshapedterm_bd_slicedterm_bd_shifteds&&&&&&&& r_relative_shift5Gemma3nAudioRelativePositionEmbedding._relative_shiftSs 0!3F/0  *=:T:T U *11  q$89   * q!:':: :  )00     rch<V^8dQhRS[PRS[PRS[P/#)rquerieskeysrIrJ)rrs"rr r!s3R)R)R)R)RXXR)rc `VPwr4rVpVPwrp\P!\P!VPVP )^, R4^R7p V P^,p VP WPR7p VPV 4p V P^WPVP4P^4pVP^^^^^4pVP^^^^^4p\P!VV4pVP^^^^^4pVP^^^4pVPW6WE,V4p\P!VV4pVPVVVVV 4pVPVVVVVV V 4pVV,#)rcrLrTr )rfr8rNr:r.r0rWrTr3rgr)r,squeeze transposematmulrn)r=rqrrr[r\r]r)r,_r^ pos_indicesr_sin_emb_timing_signalprojected_sin_embsin_emb queries_pkeys_p_tterm_ac q_transposed s_transposed q_reshapedterm_bd_unshifed_matmulterm_bd_unshifedrms&&& r__call__.Gemma3nAudioRelativePositionEmbedding.__call__s MM L &68(,zz$11nn IId''$*:*:):Q)> C! &++A. $ > > }}!?!  !MM*?@#++   '   %%aAq!4 >>!Q1a0))Ix0((Aq!Q7 ((Aq1 ")) #3#F #%))J "E3::       ..        ((r)r<r+rr,r.r0r)r3) __name__ __module__ __qualname____firstlineno__r'rWrnr__static_attributes____classdictcell__ __classcell__rErs@@rrr)s8II<++..`R)R)R)rrcaa]tRt^toV3RlV3RlltV3RlRltR V3RlRlltRtV3RlR ltV3R lR lt R t Vt V;t #)Gemma3nAudioAttentionc <V^8dQhRS[/#rr)rrs"rr "Gemma3nAudioAttention.__annotate__s6S6S{6Src<\S V`4WnVPPVnVPP VnVP VP,VnVPPVnVPPVn \^VPP^, 4Vn VPPVnVPP VnVPVP,VP,Vn\'V4Vn\*P,!VP 34Vn\0P2!VP VPVP ,RR7Vn\0P2!VP VPVP ,RR7Vn\0P2!VP VPVP ,RR7VnVP R,pR\*P:!R4, pWE,Vn\*P>!\*P@!VP$VP3\*PBR7^R7PDp\*P>!\*P@!VPVP$3\*PBR7VPVP,R7p\*P@!VPVP$3\*PBR7pW,V,pWn#\*PH!VP"\*PJR7Vn&R#) rFr#r%g@rt)kNg)'r&r'rr(r)r*r,conf_attention_chunk_size chunk_sizer/max_future_horizonr7r-max_past_horizon#conf_attention_invalid_logits_valueattention_invalid_logits_valueconf_attention_logit_capattention_logits_soft_cap context_sizerrelative_position_embeddingr8zeros per_dim_scaler1r2q_projk_projv_projr5_q_scaletrilonesbool_T_local_causal_valid_maskr;rP_softcap) r=rr>r?q_scale r_softplus_0lower_causal_maskupper_causal_masklocal_causal_valid_maskrEs &&*, rr'Gemma3nAudioAttention.__init__s  ==;;22((DNN: ++??"&++"J"J #At{{'N'NQR'R S KK ; ; +*.)M)M& OOd33 3d6M6M M ,QQW+X(XXt}}&67ii   dnnt}}<5  ii   dnnt}}<5  ii   dnnt}}<5  --%RVVC[( . GG GGT&&8 I  ! GG GGT__d&7&78 I##d&=&== #%'' __d// 0#  $ 7:K K )@%!?!?rzzR rcZ<V^8dQhRS[PRS[RS[RS[P/#)rx dim10_val dim11_valrIr`)rrs"rr rsA X X 88 X X X  Xrc^.VP,^,pVP^, p^V,pW$V&W4V^,&\P!V\\ V4VP 44#r)rr8rertuplerf)r=rrrridim_idx_from_endstart_idx_for_dims&&&& r _pad_dim1Gemma3nAudioAttention._pad_dim1sk aff q( 66A: 00+4'(/8'!+,vva7m8LaggVWWrcj<V^8dQhRS[PRS[S[S[3,RS[P/#)rr padding_valrI)r8r;rboolr6)rrs"rr rs6 ' ' '(-dEk(: '  'rcDVPpVR,wrEWPP,^, VP,pW`P,V, ;p^8dVPV^V4pWFVP3VR,,pVPV4#)NrNrNN)rfrrrg) r=rrrfbt num_blocks padding_len permute_dimss &&& r_convert_to_block'Gemma3nAudioAttention._convert_to_blocksRy//)A-$//A %7!; ;Kq @q![1At7%)C yy&&rcnVPpWR,pWc, V,^,p.p\V4F]p W,p W,p \R4.\V4,p \W4W&VP V\ V 4,4K_ \ P!W^,R7#)rcNrL)rfrangeslicerappendrr8stack) r=r dimensionsizesteprfdim_size num_windowswindowsi start_idxend_idxslicess &&&&& r unfold_mlx Gemma3nAudioAttention.unfold_mlx,s#4/!3 {#AI&GDk]SZ/F %i 9F  NN1U6]+ ,$xx!m44rcN<V^8dQhRS[PRS[P/#rrrIrJ)rrs"rr rAs#RXXrcVPpVPVP,^, pVPWV4pVPpVPpVP ^,pVP ^,pVP R,pVP V^WE4p VP^8d'V P^8dV P^^^^^4p V #)rcr) rrrrrrfrrrv) r=rpad_left pad_right frame_len frame_stepr[time_dim other_dims x_unfoldeds && r_extract_block_context,Gemma3nAudioAttention._extract_block_contextAs((++doo=A NN1 2%% __ WWQZ 771:WWR[ __Q9A 66A:*//A-#--aAq!,VPVP34pVRRV 13,pV#) Nrz%Shape of extracted_valid_mask_blocks z is not (z, z) after potential reshape.rLrdr r )NNN.) rrgrfr)r,rrr8 logaddexprrrrrvrr ValueErrorrNr logical_andrrr1tanhwherersoftmaxrOrPrTrwr) r=rr query_states key_states value_statesper_dim_scale_spbroadcast_shapeper_dim_scale_sp_broadcastr[q_time query_blocks key_blocks value_blocksr\original_valid_maskextracted_valid_mask_blockscondition_from_input_validitycondition_from_causalityfinal_condition_for_wherelogits probabilitiesb_dimn_dimu_dimw_dimc_dimh_dimprob_bunv_bun result_bmmcontext_vectorss &&& rrGemma3nAudioAttention.__call__Vsu{{1~-- WWSb\ >> +/== [[^++ WWSb\ >> +/== {{1~-- WWSb\ >> +/== <<(:(:C@aDMM2%5%=%=o%N"#mm36PP )//3 --l; 00< 22<@ '--a0 $e'+&A&A ' )Aq!  $ ( , , 1+11!4 B+11!48HH+11!4)//23  !+F*M*Md.?.?+ ' ' , ,     1  /556i |L$%R(9(9'::TV )+ 'a) %)+ )) % $(#@#@AV#W %'NN ) $% ! 11,K$--'$--' %vt/R/R  6==#<2FMM    -:,?,?)ueUE""2& **1aAq9AA"eUS&&q!Q15==b%OYYx/ $,, 5% )Aq!Q " *11 ??2   *!WfW*5r)rrrrrrrrr,r*rrrr)rrrr)r) rrrrr'rrrrrrrrrs@@rrrsI6S6Sp X X ' '5**xxxrrc\aa]tRtRtoRtRV3RlV3RllltR V3RlRlltRtVtV;t #) Gemma3nCumulativeGroupNormiaApplies Group Normalization cumulatively over the time dimension. This layer normalizes the input by calculating the mean and variance cumulatively over the time dimension (dim 1). The statistics are computed over all feature dimensions (specified by `feature_dims` and `num_channels`) for elements marked as valid by the optional `mask`. If a `mask` is provided (True for valid, False for invalid/padded), invalid time steps do not contribute to the statistics calculation, and their corresponding output values are zeroed out. Scale and bias, if enabled, are applied per-channel (last dimension). This behavior is similar to JAX's `GroupNormalization` with `num_groups=1` and `cumulative=True`. c H<V^8dQhRS[RS[S[,RS[RS[RS[/#)r num_channels feature_dimseps use_scaleuse_bias)rarr6r)rrs"rr 'Gemma3nCumulativeGroupNorm.__annotate__sLNNN  N  NNNrc <\SV`4Wn\V4VnW0nW@nWPnVP 'd\P!V4Vn MRVn VP'd\P!V4Vn MRVn \\^^\VP4,^,44VnR#N)r&r'rrrrrrr8rweightrr$rrreduction_axes)r=rrrrrrEs&&&&&&rr'#Gemma3nCumulativeGroupNorm.__init__s (!,/" >>>'',/DKDK ===.DIDI$E!QT5F5F1G-G!-K$LMrcx<V^8dQhRS[PRS[S[P,RS[P/#r)r8r;r)rrs"rr rs8e0e0"((e0(288*<e0e0rc fVPVP3,pVPR,V8wd#\RVPR, RV 24hVeyVPVPR,8wd-\RVP RVPR, 24hVP\ P 8wd \R4hVPp\ PpVPV4pVeER \V4,pVPVPV,4PV4pM%\ P!V4PV4pWh,p \ P!WPR R 7p \ P!V ^R 7p \ P!WPR R 7p \ P!V ^R 7p \ P !V ^R4pW, pWo, ^,p\ P!VV,VPR R 7p\ P!V^R 7pVV, pWo, \ P"!VVP$,4,pVP&'doVP(eaVP(PV4p^.VP*^, ,VP.,pVVPV4,pVP,'doVP.eaVP.PV4p^.VP*^, ,VP.,pVVPV4,pVV,pVPV4#) a8Applies cumulative group norm, optionally using a mask. Args: x: Input tensor, shape [B, T, *feature_dims, C]. mask: Optional boolean mask, shape [B, T]. True indicates a valid (non-padded) time step. If None, all time steps are considered valid. Returns: Normalized tensor with the same shape as x. rzInput tensor shape suffix z> does not match expected suffix (feature_dims + num_channels) Nrz Mask shape z( must match input Batch/Time dimensions zMask must be a boolean tensor.T)rMkeepdimsrLrc)rrrfrrTr8r TypeErrorrPrOrrg ones_likesumrcumsumcliprsqrtrrrrrr$)r=rrexpected_input_suffix input_dtype calc_dtypex_calcmask_suffix_shape mask_calcx_masked_for_sumsum_values_at_tcum_sum_valueselements_in_group_at_tcum_count_elementssafe_cum_count_elementscum_meansquared_diff_from_meansum_sq_diff_at_tcum_sum_sq_diff cum_variance normalized_xscalescale_view_shaper$bias_view_shape final_outputs&&& rr#Gemma3nCumulativeGroupNorm.__call__s !% 1 1T5F5F4H H 772;/ /,QWWR[M:99N8OQ   zzQWWR[( !$**-UVWV]V]^`VaUbczzRWW$ @AAgg ZZ *%   $s+@'A A  TZZ2C%CDKKJWI V,33J?I"-&& #6#6 ?;"$ //$"  YY'=AF"$''*>>dkk5KK&&z2E !saffqj1T5F5F4GG '%--8H*IIL ===TYY299##J/D cQVVaZ0D4E4E3FFO'$,,*GGL$i/ "";//r)r$rrrrrrr)gMbP?TFr) rrrr__doc__r'rrrrrs@@rrrs) NNBe0e0e0rrcTaa]tRtRtoRV3RlV3RllltV3RlRltRtVtV;t#)Gemma3nAudioSSCPConvBlockikc P<V^8dQhRS[RS[RS[RS[S[S[S[S[3,/#)ridxinput_freq_dimrmanual_padding)rarr)rrs"rr &Gemma3nAudioSSCPConvBlock.__annotate__lsA/ / / /  / c3S01 / rc x<\SV`4W0nW@nV^8Xd^M#VPPV^, ,pVPPV,pVPP V,wrVPP V,wr\P!VVV V 3W3RRR7Vn VVP^,,VP^,,p W, V ,^,p\VV3VPPRRR7Vn R#)rF) in_channels out_channels kernel_sizestriderr$T)rrrrrNr ) r&r'rrBsscp_conv_channel_sizesscp_conv_kernel_sizesscp_conv_stride_sizer1Conv2dconvr sscp_conv_epsnorm)r=r@rArrBr>r?rErFkernel_hkernel_wstride_hstride_w f_in_padded f_out_convrEs&&&&&*, rr'"Gemma3nAudioSSCPConvBlock.__init__ls  ,!8a)K)KCRSG)T {{99#> ![[>>sC![[>>sCII#%'   !!!$ %!!!$ %  ",9A= .%$ ))   rcN<V^8dQhRS[PRS[P/#rrJ)rrs"rr rCs#//"((/rxx/rc(\P!V\VPVP44pVP VP ^^^^44pVPV4pVP ^^^^4p\P!V4#r) r8rerrBrfrMrvrOr1relu)r=raudio_encodings_paddedaudio_encodings_convx_normedaudio_encodings_normeds&& rr"Gemma3nAudioSSCPConvBlock.__call__s "$ -d.A.A177K"  $yy)?)I)I!QPQST)UV9912!)!3!3Aq!Q!?ww-..r)rrMrBrO))rrrr rrrrr'rrrrrs@@rr>r>ks/ / b///rr>cPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#)#Gemma3nAudioSubSampleConvProjectionic <V^8dQhRS[/#rr)rrs"rr 0Gemma3nAudioSubSampleConvProjection.__annotate__s? ? {? rc<\SV`4WnVPp.p.p\ ^4FpVP V,wrVP V,wr^p V^, p ^p^pVVV V 3pVPV4WN,V,pVV , V ,^,pVPV4TpK \^VPVV^,R7Vn \^V^,VV^,R7Vn VPR,pVR,pVV,Vn \P!VPVPPRR7VnR#)r)r@rArrBFr#Nr )r&r'rinput_feat_sizerrJrKrr>conv_0conv_1rIinput_proj_in_featuresr1r2r*input_proj_linear)r=rr>r?current_f_for_block_inputcalculated_block_paddingcalculated_f_out_dimsrrPrQrRrS pad_t_top pad_t_bottom pad_f_left pad_f_rightmanual_padding_tuplerTf_out_after_conv final_c_out final_f_outrEs&&*, rr',Gemma3nAudioSubSampleConvProjection.__init__s    " " "$& "qA!'!=!=a!@ H!'!=!=a!@ H I#a>!Q#7 KK0 1 KKN!"%{{1aA. '//%-H''(89 r)rrfrgrhrir_rs@@rraras? ? B   rracPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#)Gemma3nAudioConformerAttentionic <V^8dQhRS[/#rr)rrs"rr +Gemma3nAudioConformerAttention.__annotate__sAA{Arc<\SV`4WnVPPVPP,pVPPV3VnVPPVn\P!VPP4Vn \VPP4Vn \V4Vn\P !VP VPPRR7Vn\VPP4VnR#)Fr#N)r&r'rr*r( post_in_shapepost_in_featuresr8r;gradient_clipping_gradient_clippingr pre_attn_normrattnr1r2post post_norm)r=rr>r?r,rEs&&*, rr''Gemma3nAudioConformerAttention.__init__s  ;;**dkk.R.RR"kkBBHM $ 7 7"$((4;;+H+H"I+DKK,C,CD)&1 II  ! !4;;#:#:  ( (?(?@rch<V^8dQhRS[PRS[PRS[P/#rrJ)rrs"rr rs3AA"((A"((ArxxArcTp\P!WP)VP4pVPV4pVP WB4pVP wrgrVP WgW,4p VPV 4p\P!WP)VP4pW0PV4,#r) r8r#rrrrfrgrr) r=rraudio_encodings_input_to_attnaudio_encodings_normaudio_encodings_attn_outrrr)r,rxs &&& rr'Gemma3nAudioConformerAttention.__call__s()% GGA///1H1H I#11!4#'99-A#H %=$B$B!i#;#C#C )&$   II. / GGA///1H1H I,~~a/@@@r)rrrrrrrrr_rs@@rrrs#AA"AAArrcPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#) Gemma3nAudioConformerFeedForwardi(c <V^8dQhRS[/#rr)rrs"rr -Gemma3nAudioConformerFeedForward.__annotate__)sLL{Lrc<\SV`4Wn\P!VPP 4Vn\VPP4Vn \P!VPPVPP^,RR7Vn \P!VPP^,VPPRR7Vn \VPP4Vn\P!VPP4VnR#)Fr#N)r&r'rr8r;rrr r*pre_layer_normr1r2 ffw_layer_1 ffw_layer_2post_layer_normconf_residual_weight_post_layer_scaler=rr>r?rEs&&*,rr')Gemma3nAudioConformerFeedForward.__init__)s  "$((4;;+H+H"I,T[[-D-DE99 KK # #T[[%<%++O<&&r:%4%>%>q!Q%G",.FF & *$$a(*D*J*J - ) // - 7 71a @ '' 555t7N7N ..9''/2///: ; r)rrrrrrrrr_rs@@rrrEs DD:rrcPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#)Gemma3nAudioConformerBlockic <V^8dQhRS[/#rr)rrs"rr 'Gemma3nAudioConformerBlock.__annotate__s < <{ r?rxrEs&&*, rr'AudioModel.__init__sV  )LV)T&6889 9 'v .9  sAc<V^8dQhRS[PRS[PRS[S[PS[P3,/#)r audio_melrrI)r8r;r)rrs"rr rsDT-T-T-3588T- rxx! "T-rc<VPV4pVP^,p^p\\VPP 44F-pWPPP V,^,,pK/ \ P!V4V,p\ P!VRVP^,^, R7pVP^8dVVP^8XdEVR,p\ P!WrP^,VP^,34pMkVPVP8XdQVP^,^8Xd9VP^,^8wd!WGP^,8Xd VR,p\ P!W'^R7pVP^,V8wd\RVP^,V4VP^,V8dVRRV13,pMFWHP^,, p \ P!V\^V 3VP44p\VP 4F wrV !W84pK VPP"^8dEVRRRVPP"13,pVRRRVPP"13,pVP^,VP^,8wdvVP^,p VP^,p W8d7W, p \ P!V\^V 3VP44pMW8dVRRV 13,p\ P$!VR,RV4pW83#) rcN)a_maxrdrLz[Warning: Subsampled mask length %s mismatch with feature length %s after gather. Adjusting.r)Nrd).N)rrfrrrrKr8r:r#r broadcast_totake_along_axisprintrer enumeraterconf_reduction_factorr)r=rrrt_subtime_stride_productstride_pair_idxindices current_maskpadding_neededrblock target_lenmask_current_lens&&& rrAudioModel.__call__s88    %%a($S)J)J%KLO ;;#D#D_#UVW#X X  M ))E"%88'' T!5!5a!81!<     "w||q'8g&Goo..q17==3CDG   7<< /$$Q'1, a A%q))g&G)).J    a E ) m""1%  !!!$u,+AvvI6 !&););A)>!>!vv 2N+\-?-?  "$..1HA#O2 ;; , ,q 0-a1UDKK4U4U1U.UVO'+Odkk.O.O+O(OPL   a O$9$9!$< <(..q1J+11!4 ,!+!>!vv 2N+\-?-?  ".+A{ {N; ((< #:CQ,,rc/pVP4Ftwr4RV9d0\V4'dWBV&K#VP^^^^4W#&K;RV9d/\V4'dWBV&KYVP^^^4W#&KpWBV&Kv V#)z conv.weightz conv1d.weight)itemsrrv)r=weightssanitized_weightsrvs&& rsanitizeAudioModel.sanitizesMMODA!$Q''+,a(+,;;q!Q+B%( A%$Q''+,a(+,;;q!Q+?%('(!$$! r)rrr) rrrrr'rrrrrrs@@rrrs&  T-T-l!!rr)r4typingrrrmlx.corecorer8mlx.nnr1baserrrlanguager rModulerrrr>rarrrrrrrrs ))$$:l)BIIl)^vBIIvrW0W0tC/ C/LO"))Od"ARYY"AJ7ryy7:7ryy7t