Vh"ddlZddlmcmZddlmZddlmZm Z m Z ddl m Z ddl Z GddZeejj dd d fd Zd Z d d Zy)N)wrapper_set_seed)compiled_function#min_cut_rematerialization_partitionnop randomizeceZdZdZdZdZy)assert_raises_regexc ||_||_yN) exception_clsregex)selfrrs \/home/dcms/DCMS/lib/python3.12/site-packages/torch/testing/_internal/optests/aot_autograd.py__init__zassert_raises_regex.__init__ s* cyr )rs r __enter__zassert_raises_regex.__enter__s rc||jk(rGt|}tj|j|st d|jd|y|t d|jd|t d)Nz*Expected exception to match regex. regex: z , exception: Tz Expected z% to be raised, instead got exception z,Expected exception to be raised but none was)rstrresearchrAssertionError)rexc_typeexc_val tracebackmsgs r__exit__zassert_raises_regex.__exit__s t)) )g,C99TZZ-$@ MZ]Y^_aa   D..//TU]T^_a aKLLrN)__name__ __module__ __qualname__rrr rrrr r s  Mrr TFc  tj||f\D cgc]} t| tjs| !}} fd} t | t t |t} t| |} |dk(rRtjtjd|} tjtjd| }| xr|}|st| |}|s||| tyt| | |||||ycc} w)aCompares func(*args, **kwargs) in eager-mode to under AOTAutograd. Compares outputs and (if check_gradients=True) gradients produced by AOTAutograd against eager-mode PyTorch. We assume that func(*args, **kwargs) succeeds in eager-mode PyTorch. cg}t|}D]H}t|tjr|j t |8|j |Jt j|\}}|i|Sr )iter isinstancetorchTensorappendnextpytreetree_unflatten)argsreconstructed_flat_argsvc_argsc_kwargs args_spec flat_argsfuncs rfunc_no_tensorsz+aot_autograd_check..func_no_tensors6sy"$Dz 2A!U\\*'..tDz:'..q1  2 "001H)TV(x((r)dynamic partition_fnautoc|jSr  requires_gradxs rz$aot_autograd_check..G PQP_P_rc|jSr r;r=s rr?z$aot_autograd_check..Hr@r)rN) r, tree_flattenr'r(r)rrrr tree_any_only outputs_msg,_test_aot_autograd_forwards_backwards_helper)r5r.kwargsr7assert_raises_regex_fnassert_equals_fncheck_gradientstry_check_data_specializationskip_correctness_checkargr6 compiled_foutany_tensor_requires_gradany_output_requires_grad compiled_outr3r4s` @@raot_autograd_checkrR s$"..f~>Iy$ FC 3 (EC FD F )#c7AdfJ ?D 1C& #)#7#7 F_ae#f #)#7#7 F_ad#e 2O7O ' D9 % \3K @0T+ACS%'=?9 Gs DDzOutputs of the operator are different in eager-mode PyTorch vs AOTDispatcher tracing. This means the operator will have incorrect output underneath torch.compile. This could be because the operator's implementation not traceable.cpddfd }||d|rt|}||dyy)Nctj|}|Dcgc]+}t|tjs|j r|-}}t ||}tj|}d}|D]>}t|tjs||jjz }@t|tjsJ|tjj||dfScc}w)NrT) allow_unused) r,arg_tree_leavesr'r(r)r<r tree_leavessumabsautogradgrad) fr.r4rL diff_argsrNflat_outsmis rcall_forwards_backwardszM_test_aot_autograd_forwards_backwards_helper..call_forwards_backwards`s**D1 $-(SC1N&&( (q$'%%c*  $A!U\\*aeegkkm#  $ "ell+++ENN''ID'III(s C/C/FcP  |\}}tj|dDcgc]}t|tj s|!}}t d|D}td|Dr"|r  td5 |dddyd} |\}} s  ||  | |yy#t$r|rYywxYwcc}w#1swYyxYw#t$r} t| t| d} ~ wwxYw#t$r} t| || d} ~ wwxYw)Nrc38K|]}|jduywr )grad_fn.0r>s r zN_test_aot_autograd_forwards_backwards_helper..check..{sHqQYYd2Hsc3$K|]}|du ywr rres rrgzN_test_aot_autograd_forwards_backwards_helper..check..|s,QqDy,sz1does not require grad and does not have a grad_fnzGradients of the operator are different in eager-mode PyTorch vs AOTDispatcher. This means the operator will have incorrect gradients underneath torch.compile. This could be because the operator's backward is incorrectly registered or not traceable.) Exceptionr,rBr'r(r)anyall RuntimeErrortyperD)r.ignore_failureorig_out orig_gradr> tensor_argsany_non_leavesrrQ compiled_graderHrGrarMr\rKs rcheckz;_test_aot_autograd_forwards_backwards_helper..checkqsC "9!T"B Hi#)"5"5d";A">^Q*QPUP\P\B]q^ ^HKHH ,), ,' 6ij :' D9 :  C '>j$&O# m% 2 x8 *  : &+    _ :  2d1gk*1 2 *d1gcl) *sX B;C  C  C& C0 D; C C C D'C==D D%D  D%)rnTFr) r\rMr.rGrHrJrKruras `` `` ` @rrErE[s= J" * *D $u% % d4(%rrv)r(torch.utils._pytreeutils_pytreer,torch.testing._utilsrfunctorch.compilerrrmake_fxr rr testing assert_closerRrDrErrrrsb $$1YY MM4 333&+$1?h$ ?D@)r