i&6bddlZddlZddlZddlmZddlmZddlmZm Z m Z ddl Z ddl Z ddl mZddlmZgdZe ee j&ee j&fZe e j&efZee j.j0Zee j.j0ZeeZhdZed d Zed d eee j@jBfd e j.j0defdZ"ed d e j.j0de#fdZ$ed GddZ%ed de j.jLde j.jLfdZ'ed de j.jLde j.jLfdZ(y)N)Mapping) dataclass)AnyOptionalUnion) compatibility)_get_qualified_name)get_acc_ops_nameget_node_targetis_node_output_tensorFxNetAccFusionsFinderlegalize_graphstable_topological_sort> call_method call_module call_functionF)is_backward_compatiblect|tr|S|jrd|jvrd|jS|jj dd}|r|ndd|jS)Nacc_opsacc_ops.z torch._opsz torch.ops.) isinstancestr __module____name__replace)kmodules V/var/www/html/engine/venv/lib/python3.12/site-packages/torch/fx/passes/tools_common.pyr r si!S )q||3!**&&%% + #&+1QZZL99 submodulesnodereturnc$|jtvs-Jddjtzd|jz|jdk(rLt|jt sJ||j}t |dt|}t|S|jdk(r@|j}|jd|jvrd|jSt|St|jt sJ|jS) a, Given a `node` returns its target typename. For "call_method" node, return node.target which is the name of that method being called. This could potential lead to conflict but should be okay because normally it's on a tensor. For "call_function" node, return typename of node.target. For "call_module" node, return typename of the module that node.target point to. If seeing "_VariableFunctionsClass" in the target name string, it will be replaced by "torch". e.g. _VariableFunctionsClass.relu would become torch.relu. zExpect op types of z, z , but found r_base_class_originrrr) opCALLABLE_NODE_OPSjoinrtargetrgetattrtyper rrr )r"r#submod submod_typer*s r r r -s$ 77' ' *; <<dggY?WW ' ww-$++s+++DKK(f&:DLI  ,, O #kk  ,f>O>O1Ov' ( %V, $++s+++{{r!cz|jjdd}|duxrt|tjS)aChecks if the node output produces a Tensor or not. NOTE: This requires to run `ShapeProp` on the containing fx graph before calling this function. This is because it works by checking the `type` metadata on the node. This metadata is produced by the `ShapeProp`. r,N)metaget issubclasstorchTensor)r#type_s r r r Ts3 IIMM&$ 'E   @E5< C vv..zz$|'@'@@l(((&&|S5H5H'R%%c*1 4r!r$ci}t|j}|D]_}||vr |jtvrd|jvr+||jvr:|j |j j||ht|j|h}|jrs|jj}|j||jtd|jvrj|jD][}|jtvr||j vr%|j||j||jt]|jD]}|jtvrd|jvr%||j vr4|j|t!|j"|j j||_|j||jt|jrst|j |jks!|xj|j zc_@|j D]}|j ||<b|S)N tensor_meta)rAr<rBrC)rR)r:r7r'r(r0rPr<rTsetrHrCpoprUrBusersrJminrA)r=resultr7r#rQuserrVrIs r __call__zFxNetAccFusionsFinder.__call__s[/1( > 3Dv~ww// )4>>)>B>N>N!ZZ--d3f4//0$(6 ?O?L 11#66::<''  ''E(! 1 $ 77*;;$<#5#55$$--d3//((//$'E0  //Cvv%66 $0 l000  ))#.03$114::3C3CC3H1L-++$$++ #,111T **+t~~=,"4"44%++3A , 2 2F1I3{> 3@ r!r9)rrrK__doc__r3fx GraphModulerNr>rrPrNodeListrrUdictNoder_rOr!r r r `s #uxx33## F&* $9$gx'($'" $LD$uxx}}g56Dr!r gmc(tjtjtjtjtj tj tjtjtjtjtjtjtjjj j"tjjj$j"tjjj&j(tjjj*j"tjjj,j"g}t.j1|j2j4d}tj6j9}|j2j4D] }|j:D]}||xxdz cc<"t=j>}|j2j4D]}||dk(s |jA|itC|dkDr|jE}|jG|fd|<|j:D]X}||xxdzcc<||dk(s|jHdk(r |jJ|vr|jM|H|jA|ZtC|dkDrtC|j4tC|j2j4kr%tOd|Dcgc] }||dk7s |c}|j2jP|_(||_|Scc}w)a Replace the graph of the given GraphModule with one that contains the same nodes as the original, but in topologically sorted order. This is used by the merge_matmul transformation below, which disturbs the topologically sorted order of its input GraphModule, so that this order is restored before further transformation. Arguments: gm: The graph module to topologically sort. It is modified in-place. Returns: The graph module in-place sorted Warning: This topological sort is NOT stable, it will NOT preserve the original node order. If you need a stable topological sort, use stable_topological_sort instead. rc|Sr9rOxenvs r z legalize_graph..5 c!fr!r&Input graph has cycles, unable to add ))operatorrEmulsubfloordivtruedivmodleltgegteqner3opsatensym_constrain_rangedefaultsym_constrain_range_for_size _assert_asyncmsg scalar_tensor_assert_scalarrdfromkeysr;r<raGraphr[ collectionsdequeappendlenpopleft node_copyr'r* appendleft RuntimeError_codegen) rfPRIORITIZED_OPSindeg new_graphr#r^queuecurrls @r rrs2            **22 33;; $$(( $$,, %%--#O( MM"((..! ,E IJJ D $K1 K  +002E ;!  LL /1C e*q.mmo&&s,<=CII 'D $K1 KT{a77o-$++2P$$T*LL&  ' e*q. 9??c"((..114u5atPUVZP[_`P`d5a4b c  **IBH I 6bs  N N c tj|jjd}tj j }t|jjDcic]\}}|| }}}|jjD] }|jD]}||xxdz cc<"g}|jjD]&}||dk(s tj||||f(i |rutj|\}} |j| fd | <| jD]3}||xxdzcc<||dk(stj||||f5|rut|jt|jjk(s!Jd|Dcgc] }||dk7s |c}|jj|_||_|Scc}}wcc}w)aR Replace the graph of the given GraphModule with one that contains the same nodes as the original, but in topologically sorted order while preserving the original node order as much as possible. This function performs a stable topological sort where nodes appear in an order that: 1. Respects data dependencies (topological ordering) 2. Preserves the original node order when there are no dependency constraints The algorithm uses Kahn's algorithm with a priority queue: nodes with all dependencies satisfied are added to a min-heap, ordered by their original position. This ensures we always process the earliest node in the original order among ready nodes. Arguments: gm: The graph module to topologically sort. It is modified in-place. Returns: The graph module in-place sorted rrhc|Sr9rOrjs r rmz)stable_topological_sort..wrnr!ro)rdrr;r<r3rar enumerater[heapqheappushheappoprrr) rfrridxr# node_to_idr^ ready_queue_rrls @r rrHs* MM"((..! ,E I$-RXX^^#<,c4c ,J, JJ D $K1 K  46KB ;!  NN;D)94(@ AB/1C {+3&&s,<=CII FD $K1 KT{a{Z-=t,DE F  y 3rxx~~#6 6 051]4ERVK[\L\$1]0^_ 6**IBH IK,@2^s, G)* G/8G/))rrrpcollections.abcr dataclassesrtypingrrrr3torch.fxtorch.fx._compatibilityr torch.fx.noder __all__tupler4r:TensorsTensorOrTensorsrarercrYrNrNamesr(r nnModuler boolr r rbrrrOr!r rs #!'' 1-  ell#T%,,%77 8 g-.   ehhmm  S Ce, :- :e,#UXX__,-#5:XX]]##-#Le,A A$A-Ae,UU-Upe,Kuxx++K0D0DK-K\e,= 4 4=9M9M=-=r!