i$VdZddlmZddlZddlmZddlm Z Gdde jZ Gdde jZ Gdd e jZ Gd d e jZGd d e jZGdde jZGdde jZGdde jZGdde jZGdde jZGdde jZGdde jZGdde jZy)zR Shared MLX layers and NCL/NCHW wrappers. Optimized for memory layout efficiency. ) annotationsNc(eZdZdfd ZddZxZS)Lambdac0t|||_yN)super__init__fn)selfr __class__s j/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/demucs_mlx/mlx_layers.pyr zLambda.__init__s c$|j|Sr)r r xs r __call__zLambda.__call__swwqzr)r z!tp.Callable[[mx.array], mx.array]rmx.arrayreturnr__name__ __module__ __qualname__r r __classcell__r s@r rr s rrceZdZddZy)Identityc|Srrs r rzIdentity.__call__srNr)rrrrrrr rrsrrc(eZdZdfd ZddZxZS) SequentialcBt|t||_yr)rr listlayers)r r$r s r r zSequential.__init__s 6l rc8|jD] }||} |Sr)r$)r rlayers r rzSequential.__call__ s[[EaA!r)r$z nn.Modulerrrs@r r!r!s #rr!cVeZdZdZ d dfd ZddZxZS) Conv1dNCLz~ Conv1d wrapper for NCL (Batch, Channels, Length) layout. MLX Conv1d expects NLC, so we transpose inputs/outputs. c ft |tj|||||||||_yN)stridepaddingdilationgroupsbias)rr nnConv1dconv r in_channels out_channels kernel_sizer+r,r-r.r/r s r r zConv1dNCL.__init__+8 II      rcp|jddd}|j|}|jdddSNr transposer2r rys r rzConv1dNCL.__call__Bs4 KK1a  IIaL{{1a##rr;rr;r;T)r4intr5rAr6rAr+rAr,rAr-rAr.rAr/boolrrrr__doc__r rrrs@r r(r(&sq            .$rr(cVeZdZdZ d dfd ZddZxZS)ConvTranspose1dNCLzK ConvTranspose1d wrapper for NCL (Batch, Channels, Length) layout. c ft |tj|||||||||_yN)r+r,r-output_paddingr/)rr r0ConvTranspose1dr2 r r4r5r6r+r,r-rIr/r s r r zConvTranspose1dNCL.__init__N: &&   )   rcp|jddd}|j|}|jdddSr9r<r>s r rzConvTranspose1dNCL.__call__es4 KK1a  IIaL{{1a##rr;rr;rT)r4rAr5rAr6rAr+rAr,rAr-rArIrAr/rBrrCrs@r rFrFJsq            .$rrFcFeZdZdZ d dfd ZddZxZS) Conv2dNCHWzg Conv2d wrapper for NCHW (Batch, Channels, Height, Width) layout. MLX Conv2d expects NHWC. c ft |tj|||||||||_yr*)rr r0Conv2dr2r3s r r zConv2dNCHW.__init__pr7rct|jdddd}|j|}|jddddSNrr:r;r<r>s r rzConv2dNCHW.__call__s8 KK1a # IIaL{{1aA&&rr@)r4rAr5rAr.rAr/rBrrCrs@r rPrPksI     .'rrPc>eZdZ d dfd ZddZxZS)ConvTranspose2dNCHWc ft |tj|||||||||_yrH)rr r0ConvTranspose2dr2rKs r r zConvTranspose2dNCHW.__init__rLrct|jdddd}|j|}|jddddSrTr<r>s r rzConvTranspose2dNCHW.__call__s8 KK1a # IIaL{{1aA&&rrN)r4rAr5rAr/rBrrrs@r rWrWs:     .'rrWc.eZdZdZddfd ZddZxZS) GroupNormNCLz Optimized GroupNorm for NCL layout. Avoids transposing NCL -> NLC, which causes strided memory access. Performs reduction on contiguous dimensions (L) instead. ct|t||_t||_t ||_t||_|jrWtj|ftj|_ tj|ftj|_yd|_ d|_yN)dtyperr rA num_groups num_channelsfloatepsrBaffinemxonesfloat32weightzerosr/r rarbrdrer s r r zGroupNormNCL.__init__ j/ -:6l ;;''TrarArbrArdrcrerBrrCrs@r r\r\s rr\c.eZdZdZddfd ZddZxZS) GroupNormNCHWz. Optimized GroupNorm for NCHW layout. ct|t||_t||_t ||_t||_|jrWtj|ftj|_ tj|ftj|_yd|_ d|_yr^r`rks r r zGroupNormNCHW.__init__rlrcz|jd|jd}}|j}|j||||zg|jdd}tt d|j }|j |d}tj||d}||z tj||jzz} | j|j} |jrRd|gdg| j dz zz} | |jj| z|jj| z} | S)Nrr;r:Trn)rqrarsrtrurvrwrfrxryrdrerir/rzs r rzGroupNormNCHW.__call__swwqz1771:1 OOQYYq!Q!V:aggabk: U1joo./D48ffZdT:t#rxxdhh'??qww' ;;FaSEJJN33EDKK//669J9J59QQE rrrrrCrs@r rrs rrc*eZdZddfd ZddZxZS)GLUNCLc0t|||_yr)rr ro)r ror s r r zGLUNCL.__init__s  rc|tj|d|j\}}|tj|zS)Nr:)ro)rfsplitrosigmoid)r rabs r rzGLUNCL.__call__ s/xx1499-12::a=  r)r;)rorArrrs@r rrs !rrc&eZdZfdZddZxZS)GELUNCLc"t|yr)rr )r r s r r zGELUNCL.__init__s rc,tj|Sr)r0gelurs r rzGELUNCL.__call__swwqzrrrrs@r rrs rrc.eZdZdZddfd ZddZxZS)FusedGroupNormGELUzFused GroupNorm + GELU using a custom Metal kernel. Replaces the pattern: GELUNCL()(GroupNormNCL/NCHW(x)) into a single kernel launch. c.t|t||_t||_t ||_tj|ftj|_ tj|ftj|_ yr^ rr rArarbrcrdrfrgrhrirjr/r rarbrdr s r r zFusedGroupNormGELU.__init__a j/ -:gg|oRZZ@ HHl_BJJ? rcvddlm}|||j|j|j|j S)Nr;)fused_groupnorm_gelu) metal_kernelsrrir/rard)r rrs r rzFusedGroupNormGELU.__call__%s+7#At{{DIItPTPXPXYYrrrarArbrArdrcrrCrs@r rrs @Zrrc.eZdZdZddfd ZddZxZS)FusedGroupNormGLUzFused GroupNorm + GLU using a custom Metal kernel. Replaces the pattern: GLUNCL()(GroupNormNCL/NCHW(x)) into a single kernel launch. Input has 2C channels, output has C channels. c.t|t||_t||_t ||_tj|ftj|_ tj|ftj|_ yr^rrs r r zFusedGroupNormGLU.__init__0rrcvddlm}|||j|j|j|j S)Nr;)fused_groupnorm_glu)rrrir/rard)r rrs r rzFusedGroupNormGLU.__call__8s(6"1dkk499dootxxXXrrrrrCrs@r rr*s @Yrr)rD __future__rtypingtpmlx.corecorerfmlx.nnr0Modulerrr!r(rFrPrWr\rrrrrrrr rs#RYYryy !$ !$H$$B!'!'H'"))'