i6%SSKrSSKJr SSKJrJrJrJrJrJ r J r SSK r SSK J r JrJr SSKJr SSKJr SSKJr SSKJrJrJr SS KJrJr SS KJr SS KJ r SS K!J"r" \ \\ RFRH\%4r&1S kr'1Skr(\ RR\ RT\ RV\ RX\ RZ\ R\\ R^\ R`\ Rb\ Rd\ Rf\ Rh\ Rj\ Rl\ Rn\ Rp0r9\\ Rt\ 4\;S'\\ Rtr<\ Rf\ Rj1r=\<\;S'\ R^\ Rb1r>\<\;S'\ RZ1r?\<\;S'\ RR1r@\<\;S'\A"\9R55rC\<\;S'\A"5rD\<\;S'\=\>-rE\<\;S'\=\>-\@-rF\<\;S'"SS5rG\ \ RR\\ RR\ \ RR4rJS\ RRS\)4SjrKS\ RRS\)4SjrLS\ RRS\)4SjrMS\ RRS\)4SjrNS\S\)4S jrOS\ RRS\)4S!jrPS\ RRS\)4S"jrQS\ RRS\)4S#jrRS\ RRS\)4S$jrSS%\S\ RRS\)4S&jrTS%\S\ RRS\)4S'jrUS\ RRS(\S\)4S)jrVS\ RRS\)4S*jrWS\ RRS\)4S+jrXS\S\)4S,jrYS\S\)4S-jrZS\S\)4S.jr[S\ RRS\\4S/jr]S\ RRS\\4S0jr^S1\ RtS\ 4S2jr_S\ RRS\)4S3jr`S\ RRS\)4S4jraS\ RRS\)4S5jrbS\ RRS6\GS7\GS\ \)\%44S8jrcS\ RRS\)4S9jrdS\S\\\4S:jreS\ RRS\)4S;jrfS\ RRS\)4S<jrgS\ RRS\)4S=jrhS\ RRS\\\4S>jriS?\\%S\)4S@jrjSA\S\ RRS\%4SBjrkS\ RRS\\ RR4SCjrlS\ RRS\\ RR4SDjrmS\SE\%4SFjrn\ \\\\\\4roSGrpSHrq\R\R1rt\\\;SI'\R\R\R1rx\\\;SJ'\R\R\R\R\R1r~\\\;SK'\x\~-r\\\;SL'\\r\ \\\4r\RS\RSM\RSN\RSN\RS\RSN\RSN0r\\\\4\;SO'SP\S\\4SQjr\"SRSS9"STSU55rS\ RRSV\\S\\4SWjrS\ RRS\)4SXjrSY\ GRSP\S\o4SZjrS[\oS\\oS\)4S]jrS^\ GRS_\oS\\4S`jr"SaSb5r"ScSd5r"SeSf5rSg\S\4ShjrSi\Sj\S_\oS\4SkjrSl\Sm\\ RRS_\oS\4Snjr\"\R1\R15r\"\R1\A"55r\"\A"5\R15r\"\A"5\R15r\"\A"5\R15r\"\R1\R15r\"\R1\R15r\"\A"5\x5r\"\x\A"55r\"\x\x5r\"\~\A"55r\"\R1\A"55r\"\R1\A"55r\"\R1\A"55r\"\R1\A"55r\"\R1\A"55r\"\R\R\R1\A"55r\"\A"5\A"55r\"\\5r"SoSp5r"SqSr5rS\ RRSs\%S\)4StjrS\ RRSs\%4SujrSS\ RRSs\%Sv\)4SwjjrS\ RRS\\4SxjrS\ RRS\\4SyjrS\)4SzjrS\ RRS{\ \\44S|jrS\ \\44S}jrS~\ RGR\GR^S\ \\ RR\\ RR\\ RR44SjrSS\ RGR\GR^S\\S\ \\ RR\\ RR44SjjrS\ GRdS\)4SjrS\ GRdS\)4SjrS\ \\\\\4S\\S\ \\\\\44SjrS\\S\\S\\4SjrSi\S\%4SjrS\ RRS\%4SjrSm\S\%4SjrS\ RRS\%4SjrS%\S\%S\ GRdSS4SjrSS\S\ RRS\ GRdS\\S\)S\ GRd4 Sjjrg)N) dataclass)AnyDictListOptionalSetTupleUnion) VkDataTypeVkMemoryLayout VkStorageType)format_target_name)EdgeOpOverload) TensorSpec) is_bufferis_lifted_tensor_constantis_param) FakeTensorFakeTensorConverter)ExportedProgram) InputKind)TensorArgument>dequantize_affine.defaultdequantize_per_tensor.tensordequantize_per_token.defaultdequantize_per_tensor.defaultdequantize_per_channel.default$dequantize_per_channel_group.default>quantize_affine.defaultquantize_per_tensor.tensorquantize_per_token.defaultquantize_per_tensor.defaultquantize_per_channel.default_VULKAN_DTYPESFP_TINT_TQINT8_TBOOL_TALL_TNONE_TFP_INT_T FP_INT_BOOL_Tcr\rSrSrSrS\\\\44SjrS\ 4Sjr S\ S\4Sjr S\ 4S jr S\ 4S jrS rg ) DtypeSetListLz Wrapper around a list of DtypeSet with broadcasting semantics. If only one DtypeSet is provided, it applies to all positions. dtype_setscN[U[5(aUUlgU/UlgN isinstancelistvals)selfr0s _/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/executorch/backends/vulkan/utils.py__init__DtypeSetList.__init__Rs$$Z66J =GL returnc,[UR5$r2lenr6r7s r8__len__DtypeSetList.__len__W499~r;idxc|US:a([UR5S:XaURS$URU$Nrr>r7rDs r8 __getitem__DtypeSetList.__getitem__Zs4 7s499~*99Q< yy~r;c2[UR5S:H$Nrr>r@s r8is_emptyDtypeSetList.is_empty`s499~""r;c:[SUR55$)z3Returns True if any position has dtype constraints.c3># UHn[U5S:v M g7f)rN)r?).0ss r8 /DtypeSetList.any_constrained..es1y!3q6A:ys)anyr6r@s r8any_constrainedDtypeSetList.any_constrainedcs1tyy111r;r6N)__name__ __module__ __qualname____firstlineno____doc__r DtypeSetrr9intrArIboolrMrV__static_attributes__r;r8r.r.LsX  54>)A#B sx #$#22r;r.noder<cURS:wag[UR[5(ag[UR[R R 5(agg)N call_functionFT)opr4targetrtorch_ops OpOverloadrcs r8is_torch_op_noderlpsF ww/!$++~..$++uzz4455 r;ctURS:wag[URR5nU[;$NreF)rfrrgrY_DQ_OPSrc node_names r8is_dequant_noderr|s0 ww/!"4;;#7#78I  r;ctURS:wag[URR5nU[;$rn)rfrrgrY_Q_OPSrps r8 is_quant_noderus0 ww/!"4;;#7#78I  r;clURS:wag[URR5nSU;$)NreFchoose_qparamsrfrrgrYrps r8is_choose_qparams_noderys0 ww/!"4;;#7#78I y ((r;c[U[RR5=(a< UR[ R :H=(a [URS5$)zFCheck if a scale node is dynamically computed via a choose_qparams op.r) r4rhfxNodergoperatorgetitemryargsrks r8is_dynamic_qscalersG 4' 1 KK8++ + 1 "499Q< 0r;clURS:wag[URR5nUS:H$)NreFrrxrps r8is_dequant_per_channel_noders1 ww/!"4;;#7#78I 8 88r;clURS:wag[URR5nSU;$)NreF view_copyrxrps r8is_view_copy_noders0 ww/!"4;;#7#78I ) ##r;clURS:wag[URR5nUS:H$)NreFzlinear.defaultrxrps r8is_linear_noders1 ww/!"4;;#7#78I ( ((r;cv[U[RR5=(a URS:H$)Nget_attr)r4rhr{r|rfrks r8is_get_attr_noders% dEHHMM * Dtww*/DDr;programcHURURR;$r2)namegraph_signature!inputs_to_lifted_tensor_constantsrrcs r8 is_constantrs 99//QQ QQr;c[U5=(d/ [X5=(d [X5=(d [X5$)zD Check if the given node is a parameter within the exported program )rrrrrs r8 is_param_noders:  4 G " 4 W # 4 %W 3 r;exported_programcURURR;agURRURnX!RRR 5;$)NF)rgrinputs_to_buffersbuffers_to_mutatevalues)rcrbufs r8is_mutable_buffer_nodersX {{*::LLL  * * < .G6F:a,,6F)r4rhr{r|rr5tupleallrks r8is_tensor_collection_noderss dEHHMM * * DII$))E"D))Z %8H%-P-PGdii6FGGG r;c[U[RR5(dgSUR;ag[URS[ 5(ag[URS[ 5(d"[URS[5(a[SURS55$g)zZ Returns true if the given node produces a tensor value, or a collection of tensor values FrTc3B# UHn[U[5v M g7fr2rrs r8rS!is_tensor_node..rr) r4rhr{r|rrr5rrrks r8is_tensor_noders dEHHMM * * DII$))E"J//$))E"D))Z %8H%-P-PGdii6FGGG r;c[U[RR5(a [ U5$[U[ [ 45(a"[U5S:Xag[SU55$g)NrFc38# UHn[U5v M g7fr2)rrQns r8rS%is_tensor_arg_node..s3d>!$$d) r4rhr{r|rr5rr?rrks r8is_tensor_arg_noder sU$ &&d## D4- ( ( t9>3d333 r;cSnURHEn[U[RR5(dM.[ U5(dM@US- nMG U$)zY For a given node, return the number of argument nodes that are associated with tensors. rrG)rr4rhr{r|r)rccountarg_nodes r8num_tensor_arg_nodesrsJ EII(EHHMM22  ( # # QJE  Lr;clSUR;ag[URS[5(ag[URS[5(d"[URS[5(a<[ SURS55(a[ URS5$g)z7 Returns the number of tensors associated a given node rrrGc3B# UHn[U[5v M g7fr2rrs r8rS&num_tensors_in_node...sC2BQz!Z((2Br)rr4rr5rrr?rks r8num_tensors_in_noder#s DII$))E"J//$))E"D))Z %8H%-P-P C$))E2BC C Ctyy'( ( r; torch_dtypecFU[;a[SUS35e[U$)z3 Returns Vulkan dtype corresponding to torch dtype z%Invalid dtype for vulkan_preprocess ())r$AssertionError)rs r8get_vk_datatyper4s-.(D[MQRSTT + &&r;ch[U5(dgURRSS5nUce/n[U[5(aUR /nOE[U[ 5(d[U[5(aUVs/sHo3R PM nn[SU55$s snf)zh Returns true if the output of the given tensor node has dtype that is supported by the Vulkan backend. TrNc32# UH o[;v M g7fr2)r$)rQdtypes r8rS.output_dtypes_are_supported..RsBM5&Ms) rrgetr4rrr5rr)rcnode_val tensor_dtypesrs r8output_dtypes_are_supportedr>s $  yy}}UD)H   M(J''!( Hd # #z(E'B'B*23(Q( 3 BMB BB4sB/c2[U5(dgURHvn[U[RR 5(a[ U5(d gM@[U[[45(dM][SU55(aMv g g)zm Returns true if all the inputs to the given tensor node have dtype that is supported by the Vulkan backend. TFc38# UHn[U5v M g7fr2)rrs r8rS-input_dtypes_are_supported..dsHx!2155xr) rrr4rhr{r|rr5rrrcrs r8input_dtypes_are_supportedrUst $  II h . ..x889 4- 0 0HxHHH r;cH[U5(dg[U5(dgg)zy Returns true if all the inputs and outputs of the given tensor node have dtype that is supported by the Vulkan backend. FT)rrrks r8io_dtypes_are_supportedrjs# 't , , %d + + r; inputs_dtypesoutputs_dtypesc :[UR5GHup4Xn[U5S:XaM[U5(at[ UR S[ [45(aUR SSRnOUR SRnXe;aSSUSUSU34s $M[ U[ [45(dM[U5HIupx[U5(dMUR SRn X;dM7SSUSUSU SU34s s $ GM UR RS5n [ U [5(a9USn[U5S:a$U RU;aSSU RSU34$g [ U [ [45(ai[U 5HZup;[ U [5(dMX#n[U5S:dM1U RU;dMCSS USU RSU34s $ g ) z Check if all tensor inputs/outputs have dtypes in the allowed sets. Returns (is_valid, reason_string) for better error reporting. rrFzinput[z] dtype z not in z][z output dtype zoutput[)Tz dtypes valid) enumeraterr?rr4rr5rrrr) rcrriargallowed_dtypes arg_dtypejsub_arg sub_dtypeout_valts r8check_node_dtypesrwsDII&&) ~ ! #  #  #((5/D%=99HHUOA.44 HHUO11 .qc)H^DTUUU/dE] + +'n !'** ' U 3 9 9I 6!$QCr!HYKxGWX  -!'4iimmE"G':&&'* ~  "w}}N'JM'--@PQQ Q  GdE] + +g&DA!Z((!/!2~&*qwwn/L GA3hqwwixGW"XXX ' r;c[URS[5(a*URSR[R :H$[URS[ 5(d"[URS[5(aMURSH:n[U[5(dMUR[R :XdM: g g)z@ Returns true if a given node contains a tensor with bool dtype rTF)r4rrrrhr`r5rrc fake_tensors r8tensor_node_is_boolrs$))E"J//yy%%33$))E"D))Z %8H%-P-P99U+K+z22$$ 2, r;cV[U5(dgURSR$)zK Returns the number of dimensions of the tensor produced by the given node Nr)rrndimrks r8ndim_ofrs' ! & & 99U   r;c[U5(dgURSn[U[5(de[ UR 5S:ag[ SUR SS55$)z` Returns True if the node's tensor has all dimensions equal to 1 except for the last dimension. FrrGc3*# UH oS:Hv M g7f)rGNrb)rQdims r8rS'is_unsqueezed_vector..s5#4Cax#4sN)rrr4rr?shaper)rctensors r8is_unsqueezed_vectorrsc ! & & YYu F fj ) )) ) 6<<1 56<<#45 55r;c[U5(a[U5(agURH'n[U5(dM[U5(dM' g g)zT Returns true if the operator used to compute the given node contains a bool tensor TF)rrrrs r8op_contains_bool_tensorrsId 3D 9 9II ( # #(;H(E(E r;c[U5(a[U5(agURH'n[U5(dM[U5(dM' g g)zk Returns true if the operator used to compute the given node contains a tensor with more than 4 dimensions TF)rtensor_node_is_high_dimrrs r8op_contains_high_dim_tensorrsI d 7 = =II ( # #(?(I(I r;czSn[UR5Hup4URU5(dMUs $ U$r2)rris_non_constant_tensor_node)r7rcprimary_arg_idxrrs r8get_primary_arg_idxrs:%)O +   + +H 5 5H, r;nn_module_typenamescV^[U[[45(a[U4SjU55$[U[R R 5(dgURRSS5nUcgUR5Hup4Uup5TH nXe;dM g M g)Nc3<># UHn[UT5v M g7fr2)$node_comes_from_any_nn_module_in_set)rQrrs r8rS7node_comes_from_any_nn_module_in_set..s! RVQ 04G H HRVsFnn_module_stackT) r4r5rrrhr{r|rritems)rcrr_packedtypename partial_names ` r8rrs$u && RV    dEHHMM * *iimm$5t> >+4a :; nAA A+4a :; nCC C+4a :; n;; ;&' n== =&' n== =&' n== =+4a :< n?? ?&'  6}oFG Gr;rbN) rYrZr[r\r]r___annotations__ classmethodr r)r r4rarbr;r8r!r!zsR O'4&:&:4H%4H$4H  4H4Hr;r! buffer_limitc^[U5(de[URS[5(a URSR 5T:$[URS[ 5(d"[URS[ 5(a"[U4SjURS55$[S[URS535e)z Checks whether the tensors produced by the given node can fit within the device's GPU buffer limit, which represents the maximum number of elements that can be stored in a GPU buffer. rc3H># UHoR5T:v M g7fr2)numel)rQrr8s r8rS&within_buffer_limit..sF5E779|+5Es"z!Cannot get numel for val of type ) rr4rrr;r5rr RuntimeErrortype)rcr8s `r8within_buffer_limitr?s $   $))E"J//yy%%',66 DIIe$d + +z$))E:JE/R/RFTYYu5EFFF>tDIIeDT?U>VWXXr;c[URS[5(a%[URSR5S:$[URS[ 5(d"[URS[ 5(aHURSH5n[U[5(dM[UR5S:dM5 g g)zL Returns true if a given node contains a tensor with more than 4 dimensions rr'TF)r4rrr?rr5rrs r8rrs$))E"J//499U#))*Q..$))E"D))Z %8H%-P-P99U+K+z22{(()A-, r;sizescf[U5S:aUSOSn[U5S:aUSOSn[U5S:aUSOSn[U5S:aUSOSnU[R:Xa US-S-nOU[R:Xa US-S-nOU[R:Xa US-S-nO}U[R :XaUS-S-nUS-S-nOXU[R :XaUS-S-nUS-S-nO3U[R:XaUS-S-nUS-S-nO[S U35eX#XE-4$) z Calculate the image extents that will be used to represent a tensor with the given sizes and memory layout in the Vulkan Delegate. rGrrr'rzUnsupported memory layout ) r?r r*r+r,r.r/r1r=)rArwidthheightchannelsbatchs r8required_image_extentsrJsK UqE"IaEe*/U2YqFJ!OuRyHE aE!HQE 333q  >66 61*" >88 8qLQ& >22 2q qLQ& >22 21*"q  >44 4q qLQ&7x@AA (* **r;extentslimitscV^^[UU4Sj[[T5555$)Nc3:># UHnTUTU:*v M g7fr2rb)rQrrKrLs r8rS$extents_are_valid..s D0C1wqzVAY&0Cs)rranger?)rKrLs``r8extents_are_validrQs Dc'l0CD DDr; tensor_sizestexture_limitsc[5n[[5H1n[X5n[ XA5(dM UR U5 M3 U$)z Given tensor sizes, determine the set of memory layouts which will prodice a texture that can fit within the specified device limits. )setr5rrJrQadd)rRrS valid_layoutsrrKs r8valid_texture_memory_layoutsrXsFEM)*(> W 5 5   f %+ r;c\\rSrSrSrS\S\4SjrS\4Sjr S\ S\ 4S jr S\ S\ 4S jr S rg ) TensorRepriz This class is a wrapper around a pair of VkStorageType and VkMemoryLayout which describes how a tensor should be represented in the Vulkan Delegate. r%r$cXlX lgr2r%r$)r7r%r$s r8r9TensorRepr.__init__s(*r;r<c<SURSURS3$)Nz TensorRepr(, rr\r@s r8__str__TensorRepr.__str__s%T../r$2D2D1EQGGr;otherc[U[5(d[$URUR:H=(a URUR:H$r2)r4rZNotImplementedr%r$r7rbs r8__eq__TensorRepr.__eq__sG%,,! !   !3!3 3 :""e&9&99 r;c.URU5(+$r2rfres r8__ne__TensorRepr.__ne__#;;u%%%r;)r$r%N)rYrZr[r\r]r r r9r r`objectr`rfrjrarbr;r8rZrZsN +]+>+HH F t &F&t&r;rZc\rSrSrSrS\\\\44SjrSr S\ S\4Sjr S\ S \SS 4S jr S\ 4S jrS \S\4SjrS \S\4SjrS \SS 4SjrSS\ S\4SjjrSS\ S\4SjjrSrg )TensorReprListi'z This class is a wrapper around a list of TensorRepr instances that automatically applies a "broadcasting" mechanism. The broadcasting mechanism allows for a single underlying TensorRepr to be used to represent multiple tensors. tensor_reprscN[U[5(aUUlgU/Ulgr2r3)r7rps r8r9TensorReprList.__init__.s$&|T::L AM r;c,[UR5$r2r>r@s r8rATensorReprList.__len__3rCr;rDr<chUS:a[U5S:XaURS$URU$rFr>rHs r8rITensorReprList.__getitem__6/ 7s4yA~99Q< 99S> !r;rNchUS:a[U5S:XaX RS'gX RU'grFr>r7rDrs r8 __setitem__TensorReprList.__setitem__<s) 7s4yA~IIaL IIcNr;cNSSRSUR55S3$)N[r_c38# UHn[U5v M g7fr2r rQtss r8rS)TensorReprList.__str__..C9ySWWyr]joinr6r@s r8r`TensorReprList.__str__B%4999tyy99:!<.os"W=V6;;=Vrc38# UHoRv M g7fr2rrs r8rSrps$ &@FKK&@rzTensorRepSet(Buffer Layouts: [z], Texture Layouts: [z]))rrr)r7buffer_layoutstexture_layoutss r8r`TensorRepSet.__str__nsY"WT=V=V"WW))$ &*&@&@$  0/??TUdTeeghhr;rbc[U[5(d[$URUR:H=(a URUR:H$r2)r4rrdrrres r8rfTensorRepSet.__eq__usH%..! !  % %)C)C C J**e.I.II r;c.URU5(+$r2rires r8rjTensorRepSet.__ne__}rlr;cf[[UR5[UR55$r2)rrUrrr@s r8copyTensorRepSet.copys* )) *C0J0J,K  r;cp[UR5S:H=(a [UR5S:H$)zP A TensorRepSet is "empty" if there are no valid representations of the tensor. rr?rrr@s r8rMTensorRepSet.is_emptys2 )) *a / XC8R8R4SWX4X r;cv[URUR-URUR-5$)zm Merge this TensorRepr with another TensorRepr, returning a new TensorRepr with the intersection of the two. rrrres r8make_intersectTensorRepSet.make_intersect9   % %(B(B B  & &)D)D D  r;cv[URUR-URUR-5$)zl Merge this TensorRepSet with another TensorRepSet, returning a new TensorRepSet with the union of the two. rres r8 make_unionTensorRepSet.make_unionrr;storagecUR[R:XaURUR;$UR[R :XaURUR ;$[SUR35e)zE Check if this TensorRepr is compatible with the given TensorRepSet. Unsupported storage type )r%r r)r$r TEXTURE_3Drr=)r7rs r8 is_compatibleTensorRepSet.is_compatiblest   =#7#7 7((D,E,EE E  ! !]%=%= =((D,F,FF F!:7;O;O:PQR Rr;c[URUR-5S:=(d% [URUR-5S:$)zU Check if this TensorRepr has any representations in common with another TensorRepr. rrres r8 any_in_commonTensorRepSet.any_in_commonsK ))E,F,FF G! K Q4--0K0KKLqP r;c2[UR5S:$rL)r?rr@s r8texture_is_validTensorRepSet.texture_is_valids4--.22r;c2[UR5S:$rL)r?rr@s r8buffer_is_validTensorRepSet.buffer_is_valids4,,-11r;c2[UR5S$rL)r5rr@s r8first_valid_buffer_layout&TensorRepSet.first_valid_buffer_layoutsD--.q11r;c2[UR5S$rL)r5rr@s r8first_valid_texture_layout'TensorRepSet.first_valid_texture_layoutsD../22r;cFUR5(a([[R[R 5$UR 5(a([[RUR55$[[RUR55$)z Pick a representation (i.e. TensorRepr) from the set of possible representations. If there are multiple valid representations, then: 1. Prefer texture storage over buffer storage 2. Pick the first available memory layout. ) rMrZr DEFAULT_STORAGEr DEFAULT_LAYOUTrrrr)rr@s r8make_tensor_reprTensorRepSet.make_tensor_reprs ==??--~/L/L   " "(($*I*I*K  m22D4R4R4TU Ur;cUR5(ag[UR5S:Xa[UR5S:Xag[UR5S:Xa[UR5S:Xagg)a) A "constrained" RepSet is one that has either: 1. A single valid texture memory layout, and no valid buffer memory layouts 2. No valid texture memory layouts, and a single valid buffer memory layout 3. Is empty In this case, it is unambiguous which representation should be used for the tensor. TrGrF)rMr?rrr@s r8is_constrainedTensorRepSet.is_constrainedsg ==?? ** +q 0S9R9R5SWX5X ** +q 0S9R9R5SWX5Xr;c,UR5(+$)z7 An "ambiguous" RepSet is one that is not constrained. )rr@s r8 is_ambiguousTensorRepSet.is_ambiguouss&&(((r;cH[5n[5nURH6nUR[R U[ R 55 M8 URH6nUR[R U[ R55 M8 X4$r2) rUrrVr!r4r r)rr)r7 buffer_set texture_setrs r8_possible_pdisTensorRepSet._possible_pdisswU e //F NN=226=;O;OP Q000F OOM33FM2<!$N 3<ii F t &F&t&  $    SZ SD S > d 3223V*V..)d) 'M 2'E.ETE$4&.#. .r;r tensor_reprc&UR[R:Xa[UR1[ 55$UR[R :Xa[[ 5UR15$[SUR35e)zN Given a TensorRepr, return a TensorRepSet that contains only that TensorRepr r)r%r r)rr$rUrr=)rs r8make_tensor_repsetrDsv=#7#77[667??  ! !]%=%= =CEK$=$=#>??6{7O7O6PQRRr; tensor_val tensor_repsetcH[5nURH<n[URU5n[ XR5(dM+UR U5 M> [ UR5S:a[UR[55$[URU5$)a `tensor_val` represents an actual tensor participating in some operator computation. `tensor_repset` represents the set of valid tensor representations that may be used for that tensor that is supported by the op implementation. `texture_limits` represents the maximum texture sizes that is supported by the GPU. Given the above, return a new TensorRepSet that contains only texture layouts that can be used to produce a valid image texture for the given tensor (i.e. fits within texture limits). r') rUrrJrrQrVr?rr)rrrSrr$rKs r8filter_invalid_reprsrPs" E&<< ()9)9=I W 5 5 ! % %m 4=  :  q M>>FF  ::>N ESDT r;c,[UR5$r2r>r@s r8rATensorRepSetList.__len__rCr;rDr<chUS:a[U5S:XaURS$URU$rFr>rHs r8rITensorRepSetList.__getitem__rwr;rNc|US:a([UR5S:XaX RS'gX RU'grFr>rys r8rzTensorRepSetList.__setitem__s. 7s499~*IIaL IIcNr;cNSSRSUR55S3$)Nr}r_c38# UHn[U5v M g7fr2rrs r8rS+TensorRepSetList.__str__..rrrrr@s r8r`TensorRepSetList.__str__rr;c8URRU5$r2rrs r8rTensorRepSetList.appendsyy$$r;cn[UR5S:Xag[SUR55$)NrTc3@# UHoR5v M g7fr2)rM)rQrs r8rS0TensorRepSetList.any_is_empty..sGYk''))Ys)r?r6rUr@s r8 any_is_emptyTensorRepSetList.any_is_emptys* tyy>Q GTYYGGGr;rX)rYrZr[r\r]r rrr9rAr_rIrzr r`rr`rrarbr;r8rrs  lD,>>? "s"|" !s!!$! ==%,%4%HdHr;rc\rSrSrSrS\S\S\RRS\ 4Sjr S\ 4S jr S \ S \S\ S\ 4S jrSSjrS\4SjrS\4SjrS\4SjrS\S\ S\4SjrS\ S\4SjrS\\\44SjrSrg ) OpRepSetsia This class is responsible for representing and managing the set of valid tensor representations that may be used for all input and output tensors of an operator. It is also responsible for maintaining synchronization rules between tensors participating in the computation. Currently, three synchronization rules exist: 1. All input tensors must use the same representation (e.g. binary ops) 2. The "primary" input and output tensors must use the same representation (e.g. group norm; the output is a tuple of out, mean, rstd; out must be the same representation as the first input x, but mean and rstd may use different representations as out) 3. All output tensors must use the same representation (e.g. choose qparams) Note that "primary" input and output tensor refers to the first non-weight input tensor and the first output tensor. Note that Some operators (such as arange) do not have any tensor inputs. Currently, the above three synchronization rules are sufficient to describe the representation requirements of all ET-VK operators. This class also provides utilities to constrain the repsets; when applying the constraints, the synchronization rules will be maintained. inputs_repsetsoutputs_repsetsop_noderScX0l[U5S:aSUlOSUl[U5S:aSUlOSUlSUl[ URR 5H=upVXn[U5(dMUcM UR5(aM7XPl O URSLUl URbXRUS:waSUl [/5n[n [ UR 5HupVXn [U5(dUR[5 M0U R5(aUR[5 M\U R5(aeURXU5n URU 5 U RU 5n M [/5n [n [U5S:Xa/[!UR"SUSU5n U RU 5 Oi[ UR"S5HMup][%U [&5(de[!XUU5nU RU5 U RU5n MO [(nUR(ayUR(aU O XRnUR(aU OU SnURU5nUR+5XR'UR+5U S'UR(aBUR(aUOU n[-[U55HnUR+5X'M UR(aBUR(aUOU n[-[U 55HnUR+5X'M XlXlUR35 g)NrGFTrr)rr?sync_args_reprsync_outs_reprrrrrrMsync_primary_io_reprrALL_STORAGES_REPSETrfilter_invalid_reprs_for_argrrrrr4r NO_STORAGErrPargs_repset_listouts_repset_listassert_sync_contraints)r7rrrrSrrarg_node_repsetr common_arg_repset arg_repsetr!common_out_repsetr out_repsetprimary_repsetprimary_in_repsetprimary_out_repset common_repsets r8r9OpRepSets.__init__s ~  ""'D #'D   ! #"'D "&D  /3$T\\%6%67KA,/O%h//&''))#$ 8%)$8$8$D!    +223q7II,1),B//$W\\2KA'*J&h// ''(;<$$&& ''(;<%..0000!>>. !'' 3$5$D$DZ$P!)3.,B// w '1 , 4 U#_Q%7!   # #$5 6$GLL$78!#z22221+^ !'' 3$5$D$DZ$P!9$  $ $&&"%&:&:;  &*%8%8!>Nq>Q /==>PQN5C5H5H5J 11 2"0"5"5"7 Q    "&";";AR 3/01&3&8&8&: #2   "&";";AR 3/01&3&8&8&: #2!1 0 ##%r;r<c<SURSURS3$)NzOpRepSets(ins=z, outs=r)r r!r@s r8r`OpRepSets.__str__s& 5 56gd>S>S=TTUVVr;rrcR[U[RR5(a)[ U5(a[ UR SX5$[U[5(a#[SU55(a [XU5$URR[R:XahX RRS:XaLURRSn[U[5(de[ UR SUX5$[!SU35e)z. Helper function to call filter_invalid_reprs rc38# UHn[U5v M g7fr2)rrs r8rS9OpRepSets.filter_invalid_reprs_for_arg..s0 .6 !! $ $hrrrGzUnhandled node type )r4rhr{r|rrrr5rrrrgr}r~rr_NotImplementedError)r7rrrSrDs r8r&OpRepSets.filter_invalid_reprs_for_args h . .3H3R3R' e$k $ ' 'C0 .60 - - 6~  LL  8#3#3 3LLDUDUVWDX8X,,##A&Cc3'' ''' e$S); "$8 "CDDr;Nc fUR(a[[UR55Hn[US-[UR55HxnURUnURUnUR 5(aM8UR 5(aMOUR U5(aMgSUSUSUSU35e M UR (a[[UR55Hn[US-[UR55HxnURUnURUnUR 5(aM8UR 5(aMOUR U5(aMgSUSUSUSU35e M UR(axURURnURSnUR 5(d:UR 5(d$UR U5(d SUSU35egggg) NrGzSynced arg repsets z and z$ have incompatible packed dim info: z vs zSynced out repsets rz?Primary arg and out repsets have incompatible packed dim info: ) rrPr?r rMrrr!rr)r7rrrirj primary_arg primary_outs r8r" OpRepSets.assert_sync_contraintss    3t4456q1uc$*?*?&@AA..q1B..q1B;;==!DD  m05;_`b_ccghjgklm  B7   3t4456q1uc$*?*?&@AA..q1B..q1B;;==!DD  m05;_`b_ccghjgklm  B7  $ $//0D0DEK//2K''))+2F2F2H2H"EEtTU`Taaefqerst3I) %r;cxURR5=(d URR5$r2)r rr!r@s r8rOpRepSets.any_is_emptys-  ! ! . . 0 XD4I4I4V4V4X r;rc URU$r2)r r7rs r8get_arg_repsetOpRepSets.get_arg_repset$$Q''r;c URU$r2)r!r=s r8get_out_repsetOpRepSets.get_out_repsetr@r;arg_i source_repsetcURUnX2:XagURU5(dgUR(a&URS5R U5(dgUR U5nX@RU'UR (aW[[UR55H5nXQ:wdM URURU5URU'M7 UR(aXR:XdUR (aURSRU5URS'UR(ae[[UR55HCnUS:wdM URURURS5URU'ME UR5 g)a Attempt to constrain the repsets of the tensors participating in this operator based on an "existing" repset of an argument. The existing repset can have two sources: * A representation may have been determined for the argument already from a prior operator * The output repset of the operator which produces the argument If the existing repset of the argument is compatible with the current operator, then constrain the repsets of this operator and apply synchronization rules. This process tries to minimize the number of transition nodes that will need to be inserted by tag_memory_meta_pass.py by maintaining existing representations for as long as possible. FrT)r rrrBrrrrPr?rrr!rr")r7rDrEarg_current_repsetnarrowedrs r8try_constrain_with_arg_repset'OpRepSets.try_constrain_with_arg_repsets$"2259  .!// >>  $ $&&q)LL&44]C'/e$   3t4456:/3/D/D088B))!,7  $ $ )) )T-@-@'+'<'<(00:  ! !! $ ""s4#8#89:AAv373H3H4<URSnX!:XagURU5(dgURU5nX0RS'UR(aX[ [ UR55H6nUS:wdM URUR U5URU'M8 UR(aURURR U5URUR'UR(ax[ [ UR55HVnX@R:wdMURUR URUR5URU'MX UR5 g)z Attempt to constrain the output repsets of the tensors participating in this operator based the repset required by a downstream operator. rFT) r!rrrrPr?rrr rrr")r7rKout_current_repsetrHrs r8try_constrain_with_out_repset'OpRepSets.try_constrain_with_out_repset sm "2215  0!//@@%44_E#+a    3t44566/3/D/D088B))!,7  $ $:>:O:O$$;00:  ! !$"6"6 7 ""s4#8#89:A000373H3H4<< 11$2F2FG--a0; ##%r;c[/5n[/5n[[URR55H1nUR UnUR UR55 M3 [[UR55H1nURUnUR UR55 M3 X4$)zp For each tensor participating in the op, pick a representation for it among the possible represetntation sets. ) rorPr?rrr rrrr!)r7args_repr_listouts_repr_listrr%r's r8pick_representationsOpRepSets.pick_representations7s (+'+s4<<,,-.A..q1J  ! !*"="="? @/*4<<89A..q1J  ! !*"="="? @:--r;)r rr!rrrr)r<N)rYrZr[r\r]rrhr{r| ImageExtentsr9r r`rrrr"r`rr_r>rBrIrNr rorSrarbr;r8rrs2E&(E&*E& E& % E&NWWE'E36EHTE E:t: d ((((==)5= =~)\)d)V.eNN,J&K.r;rattrc`SUR;=(a [URSU5$)Nspec)rr)rcrVs r8has_node_spec_attrrYOs& TYY  C7499V+)rcrVvaluerXrRvs r8set_node_spec_attrr^Ss TYY   99V D$ ##E" D4- ( ( edE] + +D SZ0Gt9E * **D(!!Z0000#) !!Z0000'>tDzlKLLr; return_firstcSUR;deURSn[U[5(a[X15(a [ X15$S$[U[ [ 45(aZU(a%[USU5(a[ USU5$S$UVs/sH n[XA5(a [ XA5OSPM" sn$[S[U535es snf)NrXrz!Cannot get attr for spec of type ) rr4rrgetattrr5rr=r>)rcrVr_rXrRs r8get_node_spec_attrrbjs TYY   99V D$ ##&-d&9&9wt"CtC D4- ( ( -4T!Wd-C-C747D) M MLPQDq(8(8GA$dBDQ Q>tDzlKLLRs'C c[US5$)Nvk_storage_typerbrks r8get_node_storage_typerfxs d$5 66r;c[US5$)Nvk_memory_layoutrerks r8get_node_memory_layoutri|s d$6 77r;ct[U[[45(a[SU55$[ US5$)Nc3:# UHn[US5v M g7f)etvk_node_reprN)rYrs r8rS has_node_repr..sIDq%a)9::Dsrl)r4r5rrrYrks r8 has_node_reprrns2$u &&IDIII!$(899r; node_reprc[U[5(a4[[U55Vs/sHo!UPM nn[ USU5 g[ USU5 gs snf)Nrl)r4rorPrr^)rcrornode_repr_lists r8 set_node_reprrrsT)^,,166I$6O0PQ0P1A,0PQ4!1>B4!19=RsAch[U[[45(a [S5e[ USS5$)Nz/get_node_repr not implemented for list of nodesrlF)r4r5rr2rbrks r8 get_node_reprrts/$u &&!"STT!$(8%@@r;rcb[U[RR5(dg[ U5(d-[ U5(a[ UR S5(aUn[ U5(aUR SnUR Sn[U[RR5(deUR Sn[U[RR5(de[U[RR5(deX2U4$USS4$)a Check if the given node argument is part of a Quantize/Dequantize chain produced by the quant workflow. If so, return the source tensor that is the input to the Q/DQ chain and the quantize/dequantize nodes in the chain. Otherwise, return the argument as is and None, None )NNNrN)r4rhr{r|rrrr)r dequant_node quant_node source_args r8maybe_skip_q_dq_arg_chainrys c588== ) )s#?388A;#?#? S ! !88A;L!&&q) *ehhmm4444__Q' *ehhmm4444, 6666|33D$r;max_search_depthcrUnSn[U[RR5(dS/4$U/nURS:waX1:a[ UR 5S:XaOfUR Sn[U[RR5(dO-URU5 US- nURS:waX1:aM[U[RR5(dS/4$URS:waS/4$[U[RR5(deX$4$)z Trace through node.args[0] of a given initial node until a placeholder node is found then return it and the list of nodes traversed. If no placeholder node is found, returns None and an empty list. rN placeholderrG)r4rhr{r|rfr?rr)rcrzcur_node search_depth traverseds r8trace_args_until_placeholderrsHL h . .Rx I ++ &<+J x}}  " ==#(EHHMM22 "  ++ &<+J h . .Rx{{m#Rx h . .. .  r;rcUR[R[R4;agUR 5R 5S:=(a! UR 5R 5S:*$)Y Check if the given tensor is in the range of 4-bit quantization and is of integer type. Firrhint8uint8minitemmaxrs r8is_in_4bit_rangersS||EJJ 44 ::<   " $ A):):)<)AAr;cUR[R[R4;agUR 5R 5S:=(a! UR 5R 5S:*$)rFirrs r8is_in_8bit_rangersS||EJJ 44 ::<   $ & E6::<+<+<+>#+EEr;dimsrc[U[5(a US:aX- nU$/nUHnUS:aX1- nURU5 M U$)zn Normalize dimension indices to be non-negative and within [0, ndim). Accepts a single int or a list of ints. r)r4r_r)rr normalizedds r8normalize_dimsrsW $ !8 LD J  q5 IA! r;nchw_dimcDUS:aX- nUS:aX:deUS- U- nU$rFrb)rrwhcn_dims r8nchw_dim_to_whcn_dimr s5!| q=X_, ,qH$H Or;c8URSUR3$)Nz: )rr)rs r8get_tensor_val_strrs! r*"2"2!3 44r;c[U5(a<[URS[5(de[ URS5$[ U5(aS[URS[ [45(deSSRSURS55S3$SUR;a [U5$[URS5$)Nrr}r_c38# UHn[U5v M g7fr2)r)rQrs r8rS#get_node_val_str..sM.*sGh1!44hrr) r4rhr{r|rr5rrr )rs r8rr&s\(EHHMM**)) HtUm , ,499GhGGHJJ8}r;cURn[U[5(a$[U[5(deURnOp[U[R R 5(a<[U[R R 5(deUR5nO [U5n[U5SUS3nURHnU[U5S-- nM US- nU$)Nz = (r_z ...)) rgr4rrYrhrirjrr rrr)rcrg target_nameout_strrs r8 node_io_strr/s [[F&.))&.1111oo FEJJ11 2 2&%**"7"78888kkm &k !$'(K=:Gyy',t33 wG Nr; buffer_nameupdated_tensorcSnSnURRHnUR[R[R 4;dM3[ UR[5(dMTURRU:XdMpURnURn O Uc SUS35eX0R;d SU35eU[R :Xa[RRUSS9nX RU'g)Nzcould not find z in source program signatureF) requires_grad)r input_specskindrr) PARAMETERr4rrrrg state_dictrhnn Parameter)rrrrrinput_s r8update_program_state_dictrBs K D))55 KKI,,i.A.AB B6::~66 ;.;;D --K 6 C  %ABC ,, ,M }.MM , y"""++N%+P'5{#r;ep cur_tensoralign_to force_updatecSnURSnXc-S:waDX6U-- n[RRR USU45R 5nSnOUnU(dU(aV[ XRU5 URSn [5RU RU5URS'U$)z Align the width of the given tensor to the given alignment value and update the state dict of the program with the aligned tensor. FrrTr) rrhr functionalpad contiguousrrrrfrom_real_tensor fake_mode) rrcrrr added_padding cur_width num_paddingaligned_tensorcur_fake_tensors r8!align_width_and_update_state_dictrbsM  $Iq h"67 ,,00 K( *,  # !"ii@))E*.0AA  % %   % r;)T)r')r'F)r} dataclassesrtypingrrrrrr r rhrs !???  >-NNI(37NEJJ$9$93>?    JJ  KK!! JJ  KK!! KK!! MM:%% MM:%% MM:%% 1U[[*,-  u{{ --/h/;; ,x,ZZL JJ<n))+,x,5E\(!,/ x/22BehhmmT%((--%8% :NNO  588== T  %((-- D  $))4)CD9ehhmm99$EHHMM$d$))4)E588==ETERR R$R ? %((-- D N ((--N+:N N  4   $  C D  &ST uxx}}   ehhmm  "'''CehhmmCC.UXX]]t* %((-- D 0 ((--0 0 !0  49 0 f ehhmm  !#!(3-! 6uxx}} 6 6 %((-- D  ehhmm   EHHMMhsmS 2 o UXX]] s  EHHMMhuxx}}.E %((-- HUXX]],C  # s  S#s]# -*)3}%&&''))+C'####!!##%% 5c.155S0n%OT/-BBC&&''))1##Q##Q##Q%%q4tNC/0).)S) $>H>H>HB Yehhmm Y3 Y3 Y %((-- D +%**+n++@E|E\EdE ** .:   &&25,5,pb.b.J SJ S< SSSS!S S<!!588==!!!! !4''(>+M+M*N!."D"D!EsuM#CEN,N,N+OP$SU^-P-P,QR&su~/T/T.UV"**+n.S.S-T0<''(>+P+P*Q0,35"45 ,ce 4 -/AB "">F&(G(G'H#%P&(G(G'H#%P$n&C&C%DceL&(G(G'H#%P(.*K*K)LceT%1'''')) E &"#% ' "!< (H(HVc.c.V DUXX]]D#D$DMUXX]]M#M. MUXX]] M# MT M7 7(=2I7888N3K8:4:> >% N8R2S>A5^!;<A    8EHHMM "HUXX]]$;Xehhmm=T TU>;<  ((--  47  8EHHMM "D$7 78 FBU\\BdBFU\\FdFsDI~.ceCcN>S(3cc5:5#5 %588== %S %33ehhmm&5 55LL5 5H !! ((--! ! !  !  \\ !r;