rh1EHdZddlZddlZddlmZddlmZddlmZm Z m Z m Z ejjjZej!ddevree ee e grej!d d(dZddej&fd Zd Zdd ej&fd Zd ej.fdZdZdZedZedZdZdZdZdZ dZ!dZ"dZ#dZ$dZ%de&de&fdZ'd\Z(Z)e'e(d z Z*d!ed"e&d#e&defd$Z+d!ed"e&d#e&defd%Z,d&Z-dejfd'Z.y))z`Importing this file includes common utility methods for checking quantized tensors and modules. N)Tensor)contextmanager)TEST_WITH_TSANIS_PPCIS_MACOS IS_WINDOWSnoneqnnpackcntj|d|zz|z |dz |dz zz |z d|zzdzS)z7Computes the output shape given convolution parameters.)npfloor) input_size kernel_sizepaddingstridedilationoutput_paddings {/var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/torch/testing/_internal/common_quantized.py_conv_output_shapers\ 88Z!g+- ;{Q 1 ?&&)/0 134~3E FHI JJcJ|tj|j}|tj|j}tj||z |zj tj }tj|||}|j |}|S)zQuantizes a numpy array.)riinfominmaxroundastypeint64clip)xscale zero_pointqminqmaxdtypeqxs r _quantizer(s} |xx"" |xx"" !e)j( ) 0 0 :B T4 B 5 B Irc<|jt|z |z}|S)zDequantizes a numpy array.)rfloat)r'r"r#r!s r _dequantizer+&s 5 J &%/A Hrc~||zj|z}tj|||j|}|S)zhRequantizes a numpy array, i.e., intermediate int32 or int16 values are converted back to given type)rrr r)r! multiplierr#r$r%qtyper's r _requantizer0,s= j.   !J .B T4 ' ' .B IrFc|tjtjfvsJ|tjk(r|tjk(sJt |tj r|j }|tjk(r|rd\}}nd\}}n |rd\}}nd\}}|j}|j}|tjk(}||k(rd} d} n|rUt|| }| }||z ||z z } t| tjtjj} d} nt|d}t|d}||z ||z z } t| tjtjj} |t|| z z } t|| } t|| } t| t| gS)xCalculate the dynamic quantization parameters (scale, zero_point) according to the min and max element of the tensor)i?)i)rr4)rr,?r)torchper_tensor_affineper_tensor_symmetricqint8 isinstancernumpyrrrfinfofloat32epsrr*int) Xr& reduce_rangeqschemer$r%min_valmax_val is_symmetricr"r#s r_calculate_dynamic_qparamsrG3s u..0J0JK KK K%,,, ###!U\\" GGI    JD$"JD$ JD$JD$eegGeegGu999L' 'G8,GhGw&4$;7Erxx 3778EJ'3'G'3'Gw&4$;7Erxx 3778Ego 66JT:.JT:.J %L#j/ **rc|t|tjr|j}tj|j tj|j }}||z }tj|jdtj}tj|jdtj}t|jdD]}|j }|j } || k(r d||<d||<3t | d} t |d}| |z |z ||<t ||tjtjj||<|t!|||z z ||<t |||||<t |||||<||fS)r2rr&r5r6)r;r7rr<rrrrzerosshapefloat64rranger=r>r?r) rAr&r$r%n_levelsr"r#irDrEs r&_calculate_dynamic_per_channel_qparamsrP\sw!U\\" GGIU#''U);)?)?$Dd{H HHQWWQZrzz 2E!''!*BHH5J :##A& ' 5%%'%%' g E!HJqM'3'G'3'G')X5E!H58RXXbjj%9%=%=>E!H 558);#<.test_fnsA) +G*73 +4*6* + + + + +s *3 )rwrxs` roverride_qenginesrzs+ NrcPtjjjdk(S)Nfbgemmreryrrqengine_is_fbgemmr} >> # # * *h 66rcPtjjjdk(S)Nr reryrrrprps >> # # * *i 77rcPtjjjdk(S)Nonednnreryrrqengine_is_onednnrr~rcPtjjjdk(S)Nx86reryrrqengine_is_x86rs >> # # * *e 33rctt|j}d||<||d<|jt |}||fS)Nr)rTrMdimpermuterU)rAaxis new_axis_listys r_permute_to_axis_zerorsHquuw(MM$M! % &'A m rc |j}t|jtj|\}}tj |}t |jdD]M} tjtj|| d|| z z|| z|||| z || z|| <O|jt|} | j|SNrr5) r&rtor7r> zeros_likerMsizeclamprrrU) rAper_channel_scaleper_channel_zero_pointr quant_min quant_maxr&permute_axis_listr^rOouts r+_fake_quantize_per_channel_affine_referencers GGE0emm1DdKA   1 C 1668A; x++ekk!A$#8I!8L2L*M(++,-.7DF\]^F_`ctuvcwxAx ++e-. /C 66%=rc|j}t|jtj|\}}tj |} t |jdD],} tj|| d|| z z|| z| | <.| jt|} | |k\| |kz} tj |} || | | <| j|Sr) r&rrr7r>rrMrrrrU) dYrArrrrrr&rXqrOmaskr^s r0_fake_quantize_per_channel_affine_grad_referencers GGE0emm1DdKA   ! B 1668A; ] AaDC*;A*>$>?BXYZB[[\1] E+, -B )Oi 0D   2 C4CI 66%=rct|tjstj|}n|j j }|j tj|tjS)Ndevicer&) r;r7rtensordetachclonerrr>)rArs r to_tensorrsQ a & LLO HHJ    44u||F+5==4 AArnreturncd|zdz SNr ry)rs r_n_onesrs Fa<r)r r!ebitsmbitsc|jtjk(sJd|z|zdksJt|dz }t||z}d||zz}tt|z dz }dt||z zt|dzd|zz z}dd|z z}t |z t|z zdz} | tz} tj | tjjtj} |jtj}|dz} || z }|jtj}||k\} tjtj| ||k}tjtj| |}|| z}|jtj}|| z}|jtj}|jtj}|t|z z dz}|t z tz|z}||z }||z }|t|z z }|jtj}tj||tj}tj |||}tj |||}| tt"z|z |z z }|jtj}||z}||z}|jtjS)aJConvert FP32 numbers to sub-byte floating point numbers with the given number of exponent and mantissa bits. Input: torch.Tensor of dtype torch.float Output: torch.Tensor of dtype torch.uint8, where the bit encoding is stored in the least significant bits. e.g. fp4: bits 0-3 empty and bits 4-7 in fp4_e2m1 encoding fp6: bits 0-1 empty and bits 2-7 in fp6_e2m3 or fp6_e3m2 encoding Note: there are no special values (NaN, inf) support in this code. Values outside the representable range of Floatx after rounding are clamped to the maximum Floatx magnitude (sign is preserved). Code below is an adaptation of https://fburl.com/code/ciwofcg4 Background 1: last answer in https://stackoverflow.com/q/8981913 Background 2: Computer Organization and Design, RISC-V edition, Chapter 3.5 r rr rIl)r&r7r*r MBITS_F32 F32_EXP_BIASrint32viewr> logical_and logical_not logical_orruint8 full_likewhere EBITS_F32)r!rrexp_biasmax_int sign_mask magic_adder max_normal min_normal denorm_expdenorm_mask_intdenorm_mask_floatsign saturate_mask denormal_mask normal_mask denormal_xnormal_xmant_odd val_to_addsign_lps r_f32_to_floatx_unpackedrs& 77ekk !! ! u9u  !! !uqy!Heem$Geem$I)e+a/0Kwu~01WUQY5G1e85TUJq8|$J  u    !I-O _EKKHMM  u{{A z>D DA u{{AOM%%e&7&7 &FJWM##E$4$4]M$RSK&&J-J/!Ju{{+J vvekk"HY./14Hl*y8KGJ H HI-.H{{5;;'H 7%++6A M:q1A K1-Ay9,u4u<=Gjj%G  !G G A 44 rc|jtjk(sJd|z|zdksJd||zz}t|dz }t|}||z}||z }|dk(}tj|dkD||z dk(} ||z } | |z t z} | j tjtz} ||zj tj} | t|z z} | | z}d||<d|z t z}|dk(r||z tz|| <nrt|D]M}td|zd|dzzD]3}||z }|d|zz |tz|z z} ||z tz} | | z| | |k(<5Otj| | |}|j tjt|z tz|z z}||z}|jtjS)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 r rr)r&r7rrrrrrrrMrrrr*)r!rrrr mantissa_maskrx_pos zero_maskr exp_biased_lpexp_biased_f32mantissa_lp_int32 mantissa_f32resultdenormal_exp_biasedrO mantissa_cmp left_shiftsign_f32s r_floatx_unpacked_to_f32res 77ekk !! ! u9u  !! !eem$Iuqy!HENM)mG KE  I %%uqyUe^4IKMUNM"X- rs %]]^^--??&! ""sFNHj+Y'Zi('(J*.D  122887