^jkfdZddlZddlZddlmZddlmZddlmZm Z m Z m Z m Z e ejjj Zdevree ee e grej'de reDcgc] }|dvs| c}Z d5dZddej*fd Zd Zdd ej*fd Zd ej2fdZdZdZedZedZdZdZ dZ!dZ"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'Z0d6d(Z1dejfd)Z2d*Z3dejfd+Z4d,\Z5Z6d-Z7 d7d.ejd/e(d0e8fd1Z9d2ejtd3fd4Z;ycc}w)8z`Importing this file includes common utility methods for checking quantized tensors and modules. N)Tensor)contextmanager)TEST_WITH_TSANIS_PPCIS_MACOS IS_WINDOWSIS_ARM64qnnpack)fbgemmx86cntj|d|zz|z |dz |dz zz |z d|zzdzS)z7Computes the output shape given convolution parameters.)npfloor) input_size kernel_sizepaddingstridedilationoutput_paddings s/var/www/ramen.bs-engineer-server.com/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.)r float)r)r$r%r#s r _dequantizer-)s 5 J &%/A Hrc~||zj|z}tj|||j|}|S)zhRequantizes a numpy array, i.e., intermediate int32 or int16 values are converted back to given type)rrr"r )r# multiplierr%r&r'qtyper)s r _requantizer2/s= j.   !J .B T4 ' ' .B IrFc|tjtjfvrtd||tjk(r!|tjk7rtd|t |tj r|j}|tjk(r|rd\}}nd\}}n |rd\}}nd\}}|j}|j}|tjk(}||k(rd} d} n|rUt|| }| }||z ||z z } t| tjtjj} d} nt|d }t|d }||z ||z z } t| tjtjj} |t|| z z } t|| } t|| } t| t!| gS) xCalculate the dynamic quantization parameters (scale, zero_point) according to the min and max element of the tensorzFExpected qscheme to be per_tensor_affine or per_tensor_symmetric, got zr9rr?rrrrzerosshapefloat64r!ranger@rArBr) rDr(r&r'n_levelsr$r%irGrHs r&_calculate_dynamic_per_channel_qparamsrSesw!U\\" GGIU#''U);)?)?$Dd{H HHQWWQZrzz 2E!''!*BHH5J :##A& ' 5%%'%%' g E!HJqM'3'G'3'G')X5E!H58RXXbjj%9%=%=>E!H 558);#<listtuplelenr<rP_snr is_quantized dequantizenormr,log10)r#x_hatidxresnoisesignalsnrsnr_dbs rrZrZs!dE]# q6SZ  #Fs1vhdSVW\S]R^!_` `38Q=ACtAcFE#J'AA    "~~ LLN Y   E zE%L%,.. VVXF 5.C #))+ F 5&  Bs'Dc#"Ktjjj}|tjj_ d|tjj_y#|tjj_wxYwwNr9backends quantizedengine)qenginepreviouss roverride_quantized_enginernsY~~''..H&-ENN#3 *2  '(  'sABA+ B+!B  Bc#K |rtjjd|rtjjyy#|rtjjwwxYwwrg)r9_C!_set_default_mobile_cpu_allocator#_unset_default_mobile_cpu_allocator)qengine_is_qnnpacks r"override_cpu_allocator_for_qnnpackrtsR;  HH 6 6 8  HH 8 8 :   HH 8 8 : sA0$A "A0 #A--A0cfd}|S)NcltD]}t|5|i|ddd y#1swY+xYwrg)supported_qenginesrn)argskwargsrl qfunctions rtest_fnz"override_qengines..test_fnsA) +G*73 +4*6* + + + + +s *3 )rzr{s` roverride_qenginesr}s+ NrcPtjjjdk(S)Nr rhr|rrqengine_is_fbgemmr >> # # * *h 66rcPtjjjdk(S)Nr rhr|rrrsrss >> # # * *i 77rcPtjjjdk(S)Nonednnrhr|rrqengine_is_onednnrrrcPtjjjdk(S)Nr rhr|rrqengine_is_x86rs >> # # * *e 33rctt|j}d||<||d<|jt |}||fS)Nr)rWrPdimpermuterX)rDaxis new_axis_listys r_permute_to_axis_zerorsHquuw(MM$M! % &'A m rc |j}t|jtj|\}}tj |}t |jdD]M} tjtj|| d|| z z|| z|||| z || z|| <O|jt|} | j|SNrr7) r(rtor9rA zeros_likerPsizeclamprrrX) rDper_channel_scaleper_channel_zero_pointr quant_min quant_maxr(permute_axis_listrarRouts r+_fake_quantize_per_channel_affine_referencers GGE0emm1DdKA   1 C 1668A; x++ekk!A$#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(rrr9rArrPrrrrX) dYrDrrrrrr(rXqrRmaskras r0_fake_quantize_per_channel_affine_grad_referencers GGE0emm1DdKA   ! B 1668A; ] AaDC*;A*>$>?BXYZB[[\1] E+, -B )Oi 0D   2 C4CI 66%=rct|tjstj|}n|j j }|j tj|tjS)Ndevicer() r>r9rtensordetachclonerrrA)rDrs r to_tensorrsQ a & LLO HHJ    44u||F+5==4 AArnreturncd|zdz SNrr|)rs r_n_onesrs Fa<r)rr#ebitsmbitsc^|jtjk7rtd|jd|z|zdkDrtdd|z|zt |dz }t ||z}d||zz}t t |z dz }dt ||z zt |dzd|zz z}dd|z z}t |z t |z zdz} | t z} tj| tjjtj} |jtj}|dz} || z }|jtj}||k\} tjtj| ||k}tjtj| |}|| z}|jtj}|| z}|jtj}|jtj}|t |z z dz}|t z t z|z}||z }||z }|t |z z }|jtj}tj ||tj}tj"|||}tj"|||}| t t$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 z(Expected x.dtype to be torch.float, got rr%Expected 1 + ebits + mbits <= 8, got rrLl)r(r9r,r<r MBITS_F32 F32_EXP_BIASrint32viewrA 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& ww%++GyQRR5y51DQYQVEVDWXYYuqy!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|jtjk7rtd|jd|z|zdkDrtdd|z|zd||zz}t |dz }t |}||z}||z }|dk(}tj |dkD||z dk(} ||z } | |z t z} | jtjtz} ||zjtj} | 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| | |}|jtjt|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 z(Expected x.dtype to be torch.uint8, got rrrr)r(r9rr<rrrrrrrPrrrr,)r#rrrr mantissa_maskrx_pos zero_maskr exp_biased_lpexp_biased_f32mantissa_lp_int32 mantissa_f32resultdenormal_exp_biasedrR mantissa_cmp left_shiftsign_f32s r_floatx_unpacked_to_f32rqsH ww%++GyQRR5y51DQYQVEVDWXYYeem$Iuqy!HENM)mG KE  I %%uqyUe^4IKMUNM"X- ._to_mx_rceilbs !G#;; KK  JJuzz'23++* % % bo  [[ M  JJ)HKK ,FF G ++g 2W 'R  rr)r(r9rr,r<rN is_contiguousramaxabs unsqueezerrAr@ float8_e4m3fnrrrXrrWrfloat8_e8m0fnusqueeze) rrr orig_shaper F8E4M3_MAXr F4E2M1_MAXr)scale_e8m0_biasedr( final_shapes rto_mxfpr6BsH }}U^^U[[99 .CDEE}}R:%**7==/9Z[eZf g    "FGGJgoo CR$R.J68BGjj7+R0::2>Gjj'Gjj'G [[!4!4599  7  !!!! u||U\\) * !>".gw!Hw**U001//*- 7 / 5>>0JK:& BA//+.)..u/C/CD)11"5 g %%rrcudacddl}||zdk7rtd|tt||dz|}|t |kDr tdt j |j|dd|dz }t j|t j |gf}t j|\}}|j|j|S)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. rNzM must be divisible by rz5E cannot be larger than the number of possible valuesr) random ValueErrorrWrPrYr9rsamplecatsortr) EM multiple_ofr(rr9possible_valuesselected_values_s rgenerate_jagged_offsrDs ;!2;-@AA5a!e[ABO 3 PQQll6=="1Eq1u#MNOii%,,s2C DEOO4OQ   e $ ' ' //r)r)r)rr)<__doc__r?rr9r contextlibr$torch.testing._internal.common_utilsrrrrr rWrirjsupported_enginesrwanyremoverrr*r-r2r:rJrSrZrnrtr}rrsrrrrrrrCrrrrrrrrrr rrrrrstrr6rrD)qes0rrMs? %gg%..22DDE ""sFNHj+Y'Zi( '9YRGX=X"Y'(J*.D  122887