Vh6ddlmZmZmZmZddlZddlmZddlm Z ddl m Z ddl m Z mZddlmZmZdd lmZmZmZdd lmZmZmZmZmZmZmZmZmZm Z d ed e d ee!fdZ"d ed e dede!de!de!de!de!de!de#de#dee!d efdZ$d e de%ee&e!e!ffde%ee&e!e!ffdede!d e f dZ'dedede d ed!eejPe)fd"eejPe#fd#e!d efd$Z*ded%edeee+efd&e de d ed'e!ded(e%e!e,efd eee+effd)Z-ded&e de d ed ef d*Z.d+ed&e d,e#d e/fd-Z0d.eee+efd/eeee+efd+ed&e de d'e!d efd0Z1d.eee+efd/eeee+efded&e de d'e!d efd1Z2 d7d2e!d&e d3e!de d4e%e!e&eeffded5e/d(ee%e!e,efd e fd6Z3y)8)AnyCallableOptionalUnionN)get_node_type_to_io_type_map)get_new_attr_name_with_prefix)_is_activation_post_process) GraphModulemap_arg)GraphNode)NSNodeTargetTypeNSSingleResultValuesType NSSubgraph) get_arg_indices_of_inputs_to_log$get_node_first_input_and_output_typeget_node_input_qparamsget_normalized_nth_inputget_number_of_non_param_argsget_target_type_strgetattr_from_fqnNodeInputOrOutputTypeop_type_supports_shadowingreturn_first_non_observer_nodenodegmreturncd}t|drw|}|jdk(rJt|jtsJt ||j}t |r t||d}|j|jd}|S)N_node_name_to_scope call_moduler) hasattrop isinstancetargetstrrr rr name)rrfqnnode_to_use_for_fqnmodules K/home/dcms/DCMS/lib/python3.12/site-packages/torch/ao/ns/fx/graph_passes.py_maybe_get_fqnr,s Cr()# 77m #dkk3/ //%b$++6F*62&>tR&K#$$%8%=%=>qA J logger_clslogger_node_name_suffix ref_node_name model_nameref_nameref_node_target_type results_typeindex_within_arg index_of_argr(c t|j|z|} t||} |||j||| ||| | | }t|| ||jj d| |fi}|S)z Given a starting graph of prev_node -> node -> next_node This function creates a new logger_cls obj and adds it after node, resulting in prev_node -> node -> logger_obj -> next_node r!)rr'rsetattrgraph create_node)rrr.r/r0r1r2r3r4r5r6r(logger_node_name target_type logger_obj logger_nodes r+_insert_logger_after_noder?*s24 ++ &dB/K   J B *-**((8H4'SUVK r-*node_to_instrument_inputs_to_ref_node_name+node_to_instrument_outputs_to_ref_node_namect}ifd}|jjD]}|jdk(r'|j t t ||d|:||vs||vrt||}||vr+||\} } t|} | D]} t ||| } t| tk(rV| j}t|||d|j|| | tjjd| | | j<yt| t j"j$j&k(rjt)| D]Z\}}|j}t|||d|j|| | tjj|| | |j<\|j+|||j<||vs||\} } t|j||d|j|| | tj,jdd| |j<|j+|||j<"t/||}|S)z Takes the graph of gm, adds loggers to the output of each node in nodes_to_instrument. Returns a GraphModule with the new graph. c"t|fdS)Nc"|jSNr')renvs r+z8add_loggers_to_model..load_arg..kss499~r-r )arGs r+load_argz&add_loggers_to_model..load_argjsq566r-outputr _ns_logger_r5r6r()r r9nodesr#rLr rr,rtyper r'r?r NODE_INPUTvaluetorchfximmutable_collectionsimmutable_list enumerate node_copy NODE_OUTPUTr )rr@rAr.r1 new_graphrKrr(r2 ref_node_typearg_indices_to_log node_arg_idxnode_arg prev_nodearg_idxargnew_gmrGs @r+add_loggers_to_modelrcZsqIC7VA 77h    W%=dB%JHU V  > > ? ? r*CAA*T+'-&Fd%K"$6(L7b,OHH~-$' $6 -F%&) II&$)4??EE-.)5 # .HMM*X%((*H*H*W*WW-6h,?LGS(+CHH I2K ) " * - $ * ( - 8 C C I I18-9$' 3C /"Q(X'00x@C NBB*U+'-"; N!II!,88>>%&!" "DII '00x@C NmVApY 'F Mr- prev_node_cnode_agm_bgraph_cscale zero_pointdtype_cast_nameclt|jdz|}t||||jd|di|}t|jdz|} t|| ||jd| di| } |jdtj ||| tj fi|S)N _input_scale_get_attr_input_zero_point_ call_function)rr'r8r:rSquantize_per_tensorquint8) rdrerfrgrhrirjscale_node_name scale_nodezero_point_node_namezero_point_nodes r+ _insert_quantize_per_tensor_noderwsS3FKK/4QRSWXO D/5)$$OR_J8 **   D& 3))("b2FO    !! j/5<<@   r-node_cgm_anode_name_prefixnode_type_to_io_type_mapc d} d} d} d} d} d}t||||\}}t||||\}}|tjk(r|tjk(sL|tjk(r|tjk(s&|tjk(r$|tj k(rt j} n||k(r.|tjk7rt jj} n|tjk(r8|tjk(r%t|||}|t j} |\} }np|tjk(r&|tjk(rd} t j} n7td|d|jd|d|jdzt!|t"rt%||}| r+| |t'||||| ||S|j)d| |fi|S| r|j)d| || fi|S| sJ| }t+||||j)d ||fi|St!|t,rg}|D]|}t%||}| r(|j)d| |fi|}|j/|>| sJ| }t+||||j)d ||fi|}|j/|~|Std t1|d ) a Given a starting graph C (derived from graph B) of ... -> prev_node_c -> node_c -> ... And a corresponding related node_a, inserts the correct dtype cast node after prev_node_c to cast into the dtype expected by node_a, resulting in: dtype_cast / ... -> prev_node_c -> node_c -> ... For example, if node_c is an int8 op and node_a is an fp32 op, this function will insert a dequant. Ntozdtype cast from  z to z needs to be implementedrp call_methodr!ztype fz is not handled)rrFP32INT8FP16 FP32_OR_INT8rS dequantizeUNKNOWNnnIdentityrrqfloat16AssertionError format_noder$r rrwr:r8listappendrP)rerxrdryrfrgrzr.r{ dtype_cast_opdtype_cast_mod_clsdtype_cast_methoddtype_cast_method_dtypedtype_cast_scaledtype_cast_zero_pointnode_input_type_a_node_output_type_anode_input_type_c_node_output_type_cnode_a_input_qparamsnew_dtype_cast_namedtype_cast_modresultsprev_node_c_innernew_dtype_cast_nodes r+_insert_dtype_cast_after_noders6M" -Qj":.**.Rj":.** !6!;!; ;!%:%?%?? !6!;!; ;!%:%?%?? !6!;!; ;!%:%G%GG(( .. !6!>!> >"XX..2777 !6!;!; ; 6 D2  +!55M6J 3 32777 !6!;!; ; "'--0163E3E3G2H M"#1V%7%7%9$::RS T  +t$M;&&##   K &!, 4 "Q"?@P"QRV"W &-&9&9#!&(' '#23)))!3!51>B&-&9&9!'&(' '#23/ 40vd;&7%8HIIr-c |jdk(rt|jdz|}t||j}t j |r|j}t||||j|j|di|}|S|jdk(r|jdvsJd|jd|jdk(rctt||d |||}t|jdz|}|j|j|j|fi|}|Stt||d |||}t|jdz|}|j|j|j|t||d fi|}|Std |jd |jd) z+ Simple copy of node_a to graph_c. rm _shadow_copy_rnrrr}ztarget  is not implementedrrrzhandling of node z with op )r#rr'rr%rS is_tensordetachr8r:_copy_node_from_a_to_crrr)reryrfrgnode_a_copy_name node_a_obj node_a_copyarg_copys r+rrsyyJW89VW  &dFMM: ??: &#**,J& 3)) II'R1A  m #}}!   8V]]O#6 7 8  ==L (-(q94wH < o-   "-- 6==8+r;KK -(q94wH < o-   "--  3FD!DE K  2 2 45YvyykI\ ]  r- subgraph_anum_non_param_args_node_ac`g}|j}||jk7r.|j|t||d}||jk7r.|j||j d}|D]}||dur|nd}|j |d}||\} } n|j |j} } d} | t| kr3| dk(rn| dk(r|dk(rn|| | |sy| dz } | t| kr3| jD]$} | dk(rn| dk(r|dk(rn || |sy| dz } &y)z This function returns `False` if the input subgraph cannot be copied by `_insert_copy_of_subgraph_a_after_input_node_c`. This usually means that there is a corner case logic for which copy is not yet implemented. rct|tr:t||}|jdk(r|jdvS|jdk(ryyt|t t fr|D]}t|tryy)NrrrmTF)r$r rr#r%rtuple) node_a_argryarg_aels r+ _can_insertz3_can_insert_copy_of_subgraph_a.._can_insertsu j$ '2:tDExx=(||';;;Z'  T5M 2  !!"d+  !r-rTnormalize_to_only_use_kwargsF) end_node start_noderrreversenormalized_argumentsargskwargslenvalues) rryrrOcur_noderrelocal_num_non_param_args_node_anorm_args_kwargs norm_args norm_kwargscur_idx kwarg_vals r+_can_insert_copy_of_subgraph_arsw E""H j++ + X+HdA> j++ + LL MMO "")/58); % ("66 t7   '%5 "I{%+[[&--{II&!|A"AQ"F"9W#5t< qLGI&%++- I!|A"AQ"F"9d3 qLG 3"H r- input_node_cinput_node_c_2c pt|ttfsJ|jg}|j}||jk7r/t ||d}|j d|||jk7r/|d}t||||||} tdt|D]} || }| } t| d||||} | S)z* TODO(before land): real docblock rrN) r$r rrrrinsert)_insert_copy_of_node_a_after_input_node_cranger) rrrryrfrz nodes_of_ar cur_node_a cur_node_c cur_idx_ards r+-_insert_copy_of_subgraph_a_after_input_node_crs lT4L 11 1%%&J""H j++ ++HdA>!X& j++ + AJ:nj$>NJ1c*o.   *  >          r-ct|tr |jn!t|tsJ|dj|j d}||\}}n|j |j }}g} i} fd} d} | t|krB| dk(r|} n| dk(r||} n | || } | j| | dz } | t|krB|jD]-\}| dk(r|| |<n| dk(r||| |<n | | |<| dz } /t| } t|}|jdk(rqt|}t|jtsJt|j}t!||j#|j|| | |}|S|jdvsJj#|j|j| | |}|S)a Assume that node_a from graph_a has args (input, (input2)?, arg1, ...), and kwargs {kw0: kwarg0, ...} Note: input2 is optional. If it equals to None, we assume that the op has a single non-param input. If it is specified, we assume that the op has two non-param inputs. Copies the underlying values of arg1..argn and kwarg0..kwargn into gm_b, and creates the corresponding nodes in graph_c. Note: observers are ignored, so if an arg is an observer we navigate up until we find a non-observer parent. If node_a is a call_module, points the module pointed to by node_a to gm_b. Creates the copy of node_a in graph_c, with input as the first arg, and all other args and kwargs pointing to the copies of the objects in gm_b created above. An example in pictures: graph A: ======== input -------------> node_a / / / (input_2)?----------/ / / / / weight -> weight_obs / / bias ---------------- graph C (derived from B): ========================= input_node_c --> node_a_copy / / / (input_node_c_2)? / / / / weight_copy ----/ / / bias_copy ------/ rTrcFt|trt|}t|}|St|tt t jfr|Stttfr D]}t|tsJd|Stdtd)Nz/handling of Node inside list is not implementedzhandling for kwarg of type r) r$r rrintfloatrSdtyperrrrP)rarryrfrgrs r+ _copy_argz<_insert_copy_of_node_a_after_input_node_c.._copy_args c4 0d;C(dD'BCJ c5%++6 7J  D%= 1 E%EDE EJ -d9o->>QR r-rr!)rpr)r$r r9rrrrrritemsrrr#r%r&rr8r:)rrreryrfrzrrrnew_args new_kwargsrrnew_arg kwarg_namenode_a_shadows_c_namenew_mod_copy_namemod_anode_a_shadows_crgrs `` @@r+rr;s!f,%$$,---q/''22 43#!1 ;!'fmm; HJ&G C N " a<"G \n8$G ' 23G 1  C N ""-!2!2!4 I a<%1Jz " \n8%3Jz "%.y%9Jz "1 XHK9:JKDQ yyM!K9:JKDQ&----- v}}5'/".. II((J@U  yy<<<<".. IIv}}h j@dd| 3|!j <nt7| tBr{| D"cgc]}"3|"j }#}"tE| D]P\}$}"|#|$}!t;|!||d |j |||t<j>j@|$d| 3|!j <RntGd tI| d|s|r.|j|| 3|j <3|j }%|rt5%|d}!|rHt7|!t8rt5|%|d}!n*t7|!tBr|!D"cgc]}"t5|"|d}!}"tK|j |%|!||||j dz|| }&|rd}'t7|&t8r2t;|&||d|'|||t<j>j@dd| }&|&}(nkt7|&tBsJg})tE|&D]E\}*}+t;|+||d|'|||t<j>j@|*d| },|)jM|,G|)}&|&}(d}-t/|j |}|dk(r t5|%|d}-tO|&|-||||%j dz}.|.3|.j <|r|.}/t5|/|d(k7rt5|/|d}/t5|/|d|(k7rt7|(t8r(tQ||(j }0|/j |0_)n@t7|(tBsJ|(D])}1tQ||1j }0|/j |0_)+t;3|.j ||d|.j |||t<jTj@dd| 3|.j <|st;3|j ||d|j |||t<jTj@dd| 3|j <|j|| 3|j <=tW||}2|2Scc}"wcc}"w)a Creates a new GraphModule consisting of the graph of C, with the meaningful nodes of A shadowing the corresponding nodes of B. For example, Graph A: a0 -> op0_fp32 -> a1 -> op1_fp32 -> a2 Graph B: b0 -> op0_int8 -> b1 -> op1_int8 -> b2 matched_node_pairs: {'op0': (op0_fp32, op0_int8), 'op1': (op1_fp32, op1_int8)} Graph C (A shadows B): / dequant0 -> op0_fp32 -> logger_a_0 / dequant_1 -> op1_fp32 -> logger_a_1 / / b0 -------------> op0_int8 -> logger_b_0 --------------> op1_int8 -> logger_b_1 In a nutshell, this function does the following for each node pair: * copies the necessary attributes and modules from gm_a to gm_b, keeping names unique * adds a dtype cast op (dequant, quant, etc) * adds a copy of node_a in gm_b's graph * adds loggers to the outputs of node_a and node_b Nc"t|fdS)Nc"|jSrErF)renv_cs r+rHz6create_a_shadows_b..load_arg..suTYY'7r-rI)rJrs r+rKz$create_a_shadows_b..load_argsq788r-rLrz$skipping shadow loggers for node_b: z, start_node_a: z , unsupportedz, unknown dtype castz, unknown input qparamsz", unhandled logic in subgraph copy_ns_logger_b_inp_rNztype z is not handled yet _dtype_cast__ns_logger_a_inp_rrr _ns_logger_a_ _ns_logger_b_),rr rr base_op_noderrr9rOr#rLr rrprintrXr'rrrrrrrrr,rr$r r?rrQrRrrWrrPrrrgetattrr0rYr )4rryrrfrr.rr{rgrK+start_node_b_to_matched_subgraph_a_and_name)end_node_b_to_matched_subgraph_a_and_name match_namematchr subgraph_bref_node_type_aref_node_type_bnode_bnode_b_is_start_nodenode_b_is_end_noder2all_op_types_support_shadowingrnode_output_type_anode_input_type_bnode_output_type_bnode_io_types_known_a_and_brr fqn_base_a fqn_base_b prev_node_brdraprev_node_c_listr`rxdtype_cast_noder0 input_logger new_loggersdtype_cast_idxdtype_cast_node_innerdtype_cast_loggernode_c_second_non_param_argrrinput_logger_modinput_logger_innergm_crs4 @r+create_a_shadows_br sK H '#?#A gGE935/02-399;  E!& J-j.E.EtL-j.E.EtL     N 3J4I4IJ     J 1*2E2EF "**""aE 99 NN76;;q>8< =  &)TT#'PP #5# @G ##*)) >fE ##.H%%..5,V4 +2:;NvW[;\:]^())F)FFH'*?*G*GG ( /:;NvW[;\:]^())C)CC'=))41I($,>?RSY[_?`>ab,-@AVAVX\-],^_`34 *1):):68)LE&++&(D%%t) %2D";:;NvW[;\:]^()vtQ&O #K6(3' #5S$B' '#@))KK.0, #*%$&M!/48*C+ &/)"$+4??EE-.)* * +AP )/4@@@&( EN+FBAN,A1J 5 $ * 3 - & ( / 8 C C I I1?-.$. 1-(../@A#B$+6'6 /3+,H))4-)-12Ja3/$Q#/KK/1 $ 0@&++,% 0H28T1EU#;HdA#N38T1EU!,5+249J9J+K(9A(6),===2>K./6t=O=T=T/U,=E]],:K 0I*//0#$))#,88>>%&!"" 0&++,("%>&++&#KK#,88>>%&!"" &fkk"0")!2!268!DE&++ C aEF tW %D KK,Tl's "_4'_9rE)4typingrrrrrStorch.ao.ns.fx.mappingsrtorch.ao.quantization.fx.utilsrtorch.ao.quantization.observerr torch.fxr r torch.fx.graphr r ns_typesrrrutilsrrrrrrrrrrr&r,rr?dictrrcTensorrrwrsetrrboolrrrr rnr-r+rs)11 @HF)&LL     ; 8C=  - ---! -  -  ------ #- -`ll04T5c?5J0Kl26dE#s(O6K1Ll l  l  l^    u$ %  ellC'(  B\J \J \JtT$Z'(\J  \J  \J  \J\J\J#3,<(=#=>\J 4d \J@8 8 8  8  8  8 vHH H #H HV(d4j()(U4d#345((  (  (  ( (VA d4j()A U4d#345A  A   A  A  A  A XLPf f f f  f !eJ ,B&C!CD f  ff'tC5E1F,F'GHffr-