+ jkfRt^RIt^RIt^RIHt^RIHt^RIHtH t H t H t H t ] !]PPP 4tR]9d$]!] ]] ] .4'd]P'R4] 'd]Uu.uF qR-9gK VNK uptR.RltRR]P*3RltRt^^]P*3R ltR ]P23R ltR tR t]R4t]R4tRtRt Rt!Rt"Rt#Rt$Rt%Rt&Rt'RRlt(^^ut)t*](!])^, 4t+RRlt,RRlt-Rt.R R!lt/R/R"lt0R#R$lt1R%t2R&R'lt3^^ut4t5R(t6R0R)R*llt7^]PpR+3R,lt9R#uupi)1z`Importing this file includes common utility methods for checking quantized tensors and modules. N)Tensor)contextmanager)TEST_WITH_TSANIS_PPCIS_MACOS IS_WINDOWSIS_ARM64qnnpackc\P!V^V,,V, V^, V^, ,, V, 4^V,,^,#)z7Computes the output shape given convolution parameters.)npfloor) input_size kernel_sizepaddingstridedilationoutput_paddings&&&&&&i/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/testing/_internal/common_quantized.py_conv_output_shapers\ 88Z!g+- ;{Q 1 ?&&)/0 134~3E FHI JJcfVf!\P!V4PpVf!\P!V4Pp\P!W, V,4P \P 4p\P!WcV4pVP V4pV#)zQuantizes a numpy array.)r iinfominmaxroundastypeint64clip)xscale zero_pointqminqmaxdtypeqxs&&&&&& r _quantizer%sy |xx"" |xx"" !)j( ) 0 0 :B 4 B 5 B IrcLVP\4V, V,pV#)zDequantizes a numpy array.)rfloat)r$rr rs&&& r _dequantizer()s 5 J &%/A HrcW,P4V,p\P!WcV4PV4pV#)zdRequantizes a numpy array, i.e., intermediate int32 or int16 values are converted back to given type)rr rr)r multiplierr r!r"qtyper$s&&&&&& r _requantizer,/s9 .   !J .B 4 ' ' .B IrFc<V\P\P39d\RV 24hV\P8Xd$V\P8wd\RV 24h\ V\P 4'dVP4pV\P8XdV'dR^?rTMR^rTMV'd^^rTM^^rTVP4pVP4pV\P8HpWg8XdRp ^p MV'd`\Wv)4pV)pWv, WT, , p \V \P!\P4P4p ^p M\VR4p\VR4pWv, WT, , p \V \P!\P4P4p V\Wi, 4, p \WJ4p \WZ4p \V 4\!V 4.#)tCalculate the dynamic quantization parameters (scale, zero_point) according to the min and max element of the tensorzFExpected qscheme to be per_tensor_affine or per_tensor_symmetric, got zEH 58);#<;ir^)supported_qenginesre)argskwargsrc qfunctions*, rtest_fn"override_qengines..test_fns7)G*7334*6*43*333s 5 A )rqrrsf roverride_qenginesrus+ NrcP\PPPR8H#)fbgemmr_rtrrqengine_is_fbgemmrx >> # # * *h 66rcP\PPPR8H#)r r_rtrrrjrjs >> # # * *i 77rcP\PPPR8H#)onednnr_rtrrqengine_is_onednnr}ryrcP\PPPR8H#)x86r_rtrrqengine_is_x86rs >> # # * *e 33rc\\VP444p^W!&W^&VP\ V44pW23#)rNrHdimpermuterO)r<axis new_axis_listys&& r_permute_to_axis_zerorsBquuw(MM! % &'A rc VPp\VP\P4V4wr\P !V4p\ VP4^,4Fqp \P!\P!W ,RW,, ,W),,4WE4W),, W,,W&Ks VP\V44p V PV4#rr/) r#rtor1r9 zeros_likerHsizeclamprrrO) r<per_channel_scaleper_channel_zero_pointr quant_min quant_maxr#permute_axis_listrXrJouts &&&&&& r+_fake_quantize_per_channel_affine_referencers GGE0emm1DdKA   1 C 1668A; ++ekk!$#8I8L2L*M(++,-.7DF\F_`ctcwx  ++e-. /C 66%=rcVPp\VP\P4V4wr\P !V4p \ VP4^,4FBp \P!W,RW*,, ,W:,,4W&KD V P\V44p W8W8*,p \P !V4p W ,W&V PV4#r) r#rrr1r9rrHrrrrO) dYr<rrrrrr#rXqrJmaskrXs &&&&&&& r0_fake_quantize_per_channel_affine_grad_referencers GGE0emm1DdKA   ! B 1668A;  ADC*;*>$>?BXB[[\ E+, -B O 0D   2 CCI 66%=rc\V\P4'g\P!V4pMVP 4P 4pVP \P!V4\PR7#)devicer#) r6r1rtensordetachclonerrr9)r<rs&&r to_tensorrsT a & & LLO HHJ    44u||F+5==4 AArc0V^8dQhR\R\/#)nreturn)r;)formats"r __annotate__rsssrc"^V,^, #rt)rs&r_n_onesrs Fa<rcHV^8dQhR\R\R\R\/#rrebitsmbitsrrr;)rs"rrrs.uuvucu#u&urcVP\P8wd\RVP 24h^V,V,^8d\R^V,V, 24h\ V^, 4p\ W,4p^W,,p\ \ V, ^, 4p^\ V4V, ,\ V^,4^V,, ,p^^V, ,p\ V, \ V, ,^,p V \ ,p \P!V \PR7P\P4p VP\P4pVR,p W , pVP\P4pW8p \P!\P!V 4W84p\P!\P!W44pW ,pVP\P4pVV ,pVP\P4pVP\P4pV\ V, , ^,pV\ , \ ,V,pVV, pVV, pV\ V, , pVP\P4p\P !W\PR7p\P"!VVV4p\P"!VVV4pV \ \$,V, V, , pVP\P4pVV,pVV,pVP\P4#)aConvert FP32 numbers to sub-byte floating point numbers with the given number of exponent and mantissa bits. Input: torch.Tensor of dtype torch.float Output: torch.Tensor of dtype torch.uint8, where the bit encoding is stored in the least significant bits. e.g. fp4: bits 0-3 empty and bits 4-7 in fp4_e2m1 encoding fp6: bits 0-1 empty and bits 2-7 in fp6_e2m3 or fp6_e3m2 encoding Note: there are no special values (NaN, inf) support in this code. Values outside the representable range of Floatx after rounding are clamped to the maximum Floatx magnitude (sign is preserved). Code below is an adaptation of https://fburl.com/code/ciwofcg4 Background 1: last answer in https://stackoverflow.com/q/8981913 Background 2: Computer Organization and Design, RISC-V edition, Chapter 3.5 z(Expected x.dtype to be torch.float, got %Expected 1 + ebits + mbits <= 8, got rDl)r#r1r'r4r MBITS_F32 F32_EXP_BIASrint32viewr9 logical_and logical_not logical_orruint8 full_likewhere EBITS_F32)rrrexp_biasmax_int sign_mask magic_adder max_normal min_normal denorm_expdenorm_mask_intdenorm_mask_floatsign saturate_mask denormal_mask normal_mask denormal_xnormal_xmant_odd val_to_addsign_lps&&& r_f32_to_floatx_unpackedrs& ww%++GyQRR5y51DQYQVEVDWXYYuqy!Hem$Gem$I)e+a/0Kwu~01WUQY5G1e85TUJq8|$J  u    !I-O _EKKHMM  u{{A z>D A u{{AOM%%e&7&7 &FWM##E$4$4]$RSK&J-J/!Ju{{+J vvekk"HY./14Hl*y8KGJ H HI-.H{{5;;'H %++6A M:q1A K1-Ay9,u4u<=Gjj%G  !G G A 44 rcHV^8dQhR\R\R\R\/#rr)rs"rrrqs.Z$Z$vZ$cZ$#Z$&Z$rc.VP\P8wd\RVP 24h^V,V,^8d\R^V,V, 24h^W,,p\ V^, 4p\ V4pW,pW, pV^8Hp\P !V^8Wr, ^8H4p Wr, p W, \ ,p V P\P4\,p Wu,P\P4p V \V, ,p W,p^W&^V, \ ,pV^8XdW, \,W&M\V4F~p\^V,^V^,,4FVpVV, pV^V,, V\,V, ,p VV, \,p W,WV8H&KX K \P!WV4pVP\P4\V, \,V, ,pVV,pVP\P4#)aConvert sub-byte floating point numbers with the given number of exponent and mantissa bits to FP32. Input: torch.Tensor of dtype uint8, where the bit encoding is stored in the least significant bits. e.g. fp4: bits 0-3 empty and bits 4-7 in fp4_e2m1 encoding fp6: bits 0-1 empty and bits 2-7 in fp6_e2m3 or fp6_e3m2 encoding Output: torch.Tensor of dtype fp32 with the dequantized value z(Expected x.dtype to be torch.uint8, got r)r#r1rr4rrrrrrrHrrrr')rrrrr mantissa_maskrx_pos zero_maskr exp_biased_lpexp_biased_f32mantissa_lp_int32 mantissa_f32resultdenormal_exp_biasedrJ mantissa_cmp left_shiftsign_f32s&&& r_floatx_unpacked_to_f32rqs  ww%++GyQRR5y51DQYQVEVDWXYYem$Iuqy!HENMmG KE  I %%uqyU^4IKMNM"- .__annotate__bsJ!!!!! u||U\\) * !rc ,^pW, p\P!\P!V4^\P!\P!\P !V44V)VR7V,P \P44p\P!V^8HR\P!W5P \P4, 44p\P!W,RV,VR7pWW3#))rrr/r) r1risnanrceillog2rrexp2r9)rr"r#E8M0_EXPONENT_BIASdescaleexponent descale_fpdata_lps&&& r _to_mx_rceilto_mxfp.._to_mx_rceilbs !#;; KK  JJuzz'23++* % % bo  [[ M  JJ)KK ,FF G ++g2W 'R  rr)r#r1rr'r4rF is_contiguousramaxabs unsqueezerr9r8 float8_e4m3fnrrrNrfloat8_e8m0fnusqueeze) rrr orig_shaper" F8E4M3_MAXr# F4E2M1_MAXr0scale_e8m0_biasedr/ final_shapes &&& rto_mxfpr>Bs }}U^^U[[99 .CDEE}}R:%**7==/9Z[eZf g    " "FGGJoo CR$R.J68BGjj7+R0::2>Gjj'Gjj'G [[!4!4599  7  !>".g!H**U001//*- 7 / 5>>0JK:& BA//+.)..u/C/CD)11"5  %%rcudac^RIpW,^8wd\RV 24h\\W!^,V44pV\ V48d \R4h\ P !VPVRRV^, 44p\ P!V\ P !V.434p\ P!V4wrxVPV4PV4#)a Utility function for tests and benchmarks. Generates a tensor of length E, containing random values divisible by `multiple_of`, from 0 to M, in sorted order, and where the final value in the tensor is always M. Args: E (int): The length of the tensor. M (int): The maximum value in the tensor. Returns: torch.Tensor: A tensor of length E with the specified properties. NzM must be divisible by z5E cannot be larger than the number of possible valuesr) random ValueErrorrNrHrPr1rsamplecatsortr) EM multiple_ofr#rrApossible_valuesselected_values_s &&&&& rgenerate_jagged_offsrLs !2;-@AA5!e[ABO 3 PQQll6=="1Eq1u#MNOii%,,s2C DEOO4O   e $ ' ' //r)rwrr) )rMr):__doc__r7r r1r contextlibr$torch.testing._internal.common_utilsrrrrrrNr`rasupported_enginesrnanyremoverrr%r(r,r2rBrKrQrerkrurxrjr}rrrrrrrrrrrrrrrrrrrrr>rrL)qes0rrUs %gg%..22DDE ""sFNHj+Y'Z'Zi( '9Y'9GX=X""'9YJ*.D  122887