Mjz ddlZddlZddlZddlZddlmZmZmZer ddlZddl m Z neZ ddl Z ddl m cmZddl mZddlmZddlmZddlmZddlmZdd lmZdd lmZd gZej<eZ e jBjEed Z#d ejHde%dzfdZ&d ejHdedfdZ'ed ddede de%de(ddf dZ)y)N)AnyOptional TYPE_CHECKING)ShapeEnv)fx) is_sparse_any) compatibility)lazy_format_graph_code) py_sym_types)SymNode) GraphModuleinsert_deferred_runtime_assertsgraph_code_verbosenodereturncxd|jvr|jdSd|jvr|jdSy)zx Get the example value key for a node, since dynamo uses "example_value" while non-strict export uses "val. example_valuevalN)metars ]/var/www/html/ia-enhanced/venv/lib/python3.12/site-packages/torch/fx/passes/runtime_assert.py_get_example_valuer"s= $))#yy)) $)) yy sympy.Exprcft|}t|tr|jjSyN)r isinstancer rexpr)rrs r _get_sym_valr/s' T "C#|$xx}} rT)is_backward_compatiblegm shape_envnameexportc$%&'()*+,-./0123456789:;<=ddl;ddlm-ddlm)m*m%m&m'm 2m (m }ddl m 6ddlm}m}ddlm}|j&j)8j*4t,j.j14= 0. These asserts can happen unpredictably during fake tensor propagation, so we cannot conveniently insert them into the FX graph when they occur. Instead, we accumulate them in the ShapeEnv, and in this pass insert them into the graph as proper tests. This pass also deduplicates size-related computation, CSE-ing ops that produce symbolic values and/or are involved in runtime asserts. Additionally, shape calls (size/stride/storage_offset) are turned into compute on input sizes if possible, allowing intermediate tensors to be freed earlier. For example, here dynamo will DCE the cat and repeat calls: z = torch.cat([x, x], dim=0) # 2*s0 w = z.repeat(y.shape[0]) # 2*s0*s1 _w = w.shape[0] # something with _w, but not w ... # turns into -> _w0 = 2 * s0 _w = _w0 * s1 # where s0, s1 are either SymInt graph inputs, or the result of added size calls Redundant torch._check or torch.ops.aten._assert_scalar.default calls that assert the same expression, and redundant constrain_range calls are also deduplicated. Additionally, because single-symbol bound checks (e.g. u0 >= 0, u0 <= 5) accumulate information in the ShapeEnv, the ShapeEnv contains min/max bounds for each symbol, and we delete all previous calls, adding bound checks at the end of this pass. rN)_set_node_metadata_hook)_get_placeholder_expr#_has_uninterpretable_sympy_function CallMethodKey ConvertIntKey DivideByKey free_symbolsInnerTensorKeyresolve_unbacked_bindings)int_oo) OptimizedPythonReferenceAnalysisPythonReferenceAnalysis) ValueRangesz%sz$pre insert_deferred_runtime_asserts T)colored placeholderrrct|x}duxr@t|j xr'| xrtd|jDS)z If a size/stride/storage offset call on an intermediate tensor, we can try to compute the value from input shapes instead. Nc3K|]d}t|tjxrDtt|tj tj fxr|jdk7fyw)r4N)rrNodertorchTensorSizeop).0args r z\insert_deferred_runtime_asserts.._is_intermediate_tensor_sym_call..s]3(,1#6uzz8RS,FFm+,sA*A,)rrNumberanyargs)rrr(sympys r _is_intermediate_tensor_sym_callzIinsert_deferred_runtime_asserts.._is_intermediate_tensor_sym_callsf !& &St 3 sELL11 8<<  99  rrr stack_tracenn_module_stackcustomskip_valc|stjd|j} |j}|jdk(rYt |jt s!tdt|jt|d|j}|dd}|||j<|||jd<|||jd<|||jd<yy#t$rY@wxYw) Ncdt|tjjr t |S|Sr)rr8rr7r)r=s rzNinsert_deferred_runtime_asserts.._node_metadata_hook..s)/9#uxx}}/M&s+SVr call_methodzExpected str target, got rrDrErF) pytreetree_maprAtargetr;rstrAssertionErrortypegetattrrNotImplementedError)rrDrErFrG fake_argsrOval_keys r_node_metadata_hookz._node_metadata_hooks I  77m+%dkk37,7T[[8I7JK%Yq\4;;?F )!" I%+Y%7 '"  "'2DIIm $  &+:DII' (  "(DIIh  '   sBC!! C-,C-c ddlm}m}m}ddlm}ddlm}m}||vr||St|||||fr | ||S| |jDcgc] } || c}|||<||Scc}w)Nr)Integerr?Symbol) BooleanAtom)_run_sympy_handler sympy_interp) rBrYr?rZsympy.logic.boolalgr[torch.utils._sympy.interpr\r]rrA) expr_to_proxyrrYr?rZr[r\r]r=Analysis _sympy_interps rrbz6insert_deferred_runtime_asserts.._sympy_interps113N =  & & dWffkB C-> >1 :>)) D3]=# . D  d T"" EsA. rrcXt|jdk7s$|jjjfvry|j\}}t |j xrt |jxs.t |j xrt |jS)NF)lenrAfuncLessThan GreaterThanrrZr?)rlhsrhsrBs r_is_bound_expr_for_symbolzBinsert_deferred_runtime_asserts.._is_bound_expr_for_symbols tyy>Q $))ENNEDUDU3V"V99S3 -O*S%,,2O sELL ) Kjell.K rc  |D]}|j vspt|jjdk(r.add_runtime_assertss2 +B=(,,-2T"''"6"678.match_symbol8s"65<<8&v{{G<&!6v{{!CCAU\\]24%-- 3,0IIMM-,H04 >O0P(,yy}}X'>   M02xxV/LM!, M ":A}Q?OP3=9 M Ms DDcSrrsrrJz1insert_deferred_runtime_asserts..PsDrcjtjjjj fSr)r}r8r~rsym_sizeintrirsrrJz1insert_deferred_runtime_asserts..Vs/E$7$7 %  7 7 ; ;dAY%rcjtjjjj fSr)r}r8r~r sym_striderrsrrJz1insert_deferred_runtime_asserts..^s/(;(;$)IINN$=$=$A$AD!9)"rcjtjjjj fSr)r}r8r~rsym_storage_offsetr)rrsrrJz1insert_deferred_runtime_asserts..ds-E$7$7 %  A A I ID7%rcondc4jD]}|vsyy)NTF)r,)symbolr`sym_exprs rhas_new_untracked_symbolszBinsert_deferred_runtime_asserts..has_new_untracked_symbolss("*"7"7(!6#'(!runbacked_bindingsc: tdD]}|vsyy)Nz"resolved_unbacked_bindings is NoneTF)rQ)rmr`resolved_unbacked_bindingss rhas_new_unbacked_bindingszBinsert_deferred_runtime_asserts..has_new_unbacked_bindingss419,-QRR9(m3#'(!rrnzCSE node %s -> %s for expr %src|dk(r|St|dk\r8t|dr(t|dtjr |djdk(rWj t jjjj||djf|ddS|djdk(rWj t jjjj||djf|ddSj|dj||djf|ddSt|dr|djdk(rIj t jjjj|f|ddSj|dj|f|ddSt|dtjr9j t j"||djf|ddSt|dr-j t j$|ddf|ddSt|dr9j t j&||dj(f|ddSt|dr/j t*||dj,f|ddSt/d|) NrrdrrLsizestridestorage_offsetzunrecognized keypath )rerrM SequenceKeyr#r}r8r~rrridxrrKrroperatorgetitemsym_itefloordivdivisorrS inner_namerQ)rkeypathr)r*r+r-gors rrz+insert_deferred_runtime_asserts..gos"b=#'KLA- *71:} E *71:v7I7I J&qz&8')$)$7$7(- (?(?(C(C)-wqz~~(>%&%,ABK ("!" 'qz(:')$)$7$7(- (A(A(E(E)-wqz~~(>%&%,ABK ("!"$& % 1 1$+AJOOdGAJNN5K!"!( $ ( MB&qz2BB')$)$7$7(- (I(I(Q(Q)-%&%,ABK ("!"$& % 1 1'!*//D7 KWUVUW[$( F4F4FG#% % 3 3$,$4$4tWQZ^^6L!"!( $ ( MB#% % 3 3EMMD!Q< P ' $( K@#% % 3 3$,$5$5gaj>P>P7Q!"!( $ ( NC#% % 3 3$+dGAJ4I4I-J!"!( $#13H 1R"SSrrcJ| fvry t|S#t$rYywxYwr)r TypeError)rr/s rconvertz0insert_deferred_runtime_asserts..convertus5& 11#'(#&q6M((#'(s  ""rorprqz%deleting unused reified symbol for %s)NNNF)VrB(torch._export.passes._node_metadata_hookr&%torch.fx.experimental.symbolic_shapesr'r(r)r*r+r,r-r.torch.utils._sympy.numbersr/torch.utils._sympy.referencer0r1torch.utils._sympy.value_rangesr2deferred_runtime_assertscopyrrproxyGraphAppendingTracergraph_code_logrur setnodesr;rr7boolrr8rPdictrlist enumerateinserting_beforerrr9rrrrpoprO_checkr~rrrrAkwargsr|r erase_nodeusersrzr{rreplace_all_uses_withrtr?logicboolalgr[rsym_constrain_rangesym_constrain_range_for_sizeitemsry var_to_rangeis_intuppersysmaxsizelower _default_unspecified_value_rangeissubsetr}rZ)>r!r"r#r$r.r0r1r2 placeholdersfirst_non_placeholderrCrrrrtrr assert_exprr=rr hash_nodedefsrri0rvrrmin_valgemax_vallerrrar)r*r+r-r'r(rkrWr&rbrrr`r,rrrr/rrrrrBrrVs>` @@@@@@@@@@@@@@@@@@@@@@@@@@rrr6sp PP   2<66;;=M HHE XX * *5 1F 24& 92t 13M5L  # 77m #$( !    T " # rww 4 &G  dii '%G  dii   #'15(, ")hhmm")4Z")c3h$.")S#X% ")  ")  ")J&)UM69e *0&6VH#,     3+3+3+j  EU3BZ(v)4 # # 4E!a%L:O r ) $&8&>>]KQQ0]L9=0a%,,? )!&&( 31$)+$-ahhj$9DAq( !!"%,,.,,#M$5$5dB$?@{{  --55(,yytyy|dkkoof6MDL'3D'99 mK#}4CHH''-!#rww/ ++C0!%%k2 =(!-d!33X@!.Gtyy}}-@"E.*! -8> 9 ;3575%-- 3,0IIMM-,H04 >O0P(,yy}}X'>  7D - (7M(3 !.h 7 < 499==)<=% #4"9"9";^QJAwKKNHTHTT -4%-- 3,0IIMM-,H04 >O0P(,yy}}X'>   02xx "4 1&0M!,  ":A}Q?OP}^Q@_ )#''B/P55++B/99S[[1_!<%RXXv6B AACLLRP (1!))/(, m(D,0IIMM:K,L$(99==#:  =(/rxx'88GE!.}bGm!L!Q!QB!// %  = = E E$&&NrU\}o]ghjgkkl$m!"*--bGm<'.rxx'88GE!.}bGm!L!Q!QB!// %  = = E E$&&NrU\}o]ghjgkkl$m!"*--bGm<==@-004#C(_ )gr )r ) v)r %**,, e tU\\ * .JJ$$ II=t D HH   +,mH  \==c r )r )s_5K i > h# Fi A+i "h0 'Di Dh= $i #h- (i 0h: 5i =i i  i )F)*rzloggingrrtypingrrrrBrrr8torch.utils._pytreeutils_pytreerMrtorch._subclasses.meta_utilsrtorch.fx._compatibilityr torch.fx._utilsr "torch.fx.experimental.proxy_tensorr torch.fx.experimental.sym_noder torch.fx.graph_moduler __all__ getLogger__name__rt_logginggetArtifactLoggerrr7rPrrrrrrrrs //>H $$612;2- - -g!11(