EiK:<SSKJr SSKrSSKJs Jr SSKJr SSKJ r SSK J r SSK J r SSKJrJr SSKJr SS KJrJrJr SS KJrJr S rSS jrS \R84SjrS \R84SjrSrSr Sr!Sr"\!"5r#\""5r$Sr%\%"5r&Sr'g))castN)_prims) DispatchKey)autograd_not_implemented)HigherOrderOperator)CUDARngStateHelpermake_contiguous_strides_for)FakeTensorMode)disable_proxy_modes_tracingProxyTorchDispatchModetrack_tensor_tree)_device_dtypec h[SURSURSURS35e)Nz"You are trying to functionalize a z RNG operator but zE does not use Philox/counter-based RNG. Therefore, functionalizing a zo RNG operator is not supported. We are discussing the possibility of a Philox-based RNG implementation for CPU.) RuntimeErrortype)devices S/home/mande/repo/quber/.venv/lib/python3.13/site-packages/torch/_prims/rng_prims.pythrow_on_non_cudarsB  ,V[[M9KFKK=YFFLkk]Si i c[RRSU-USUS9nURU5 [ [R R RU5nURnU(aXXl Xx4HGn XIl [RRRU lX-U lX)lX9lMI g)Nz rngprims::) mutates_argsschema)torchlibrary custom_op register_fakegetattr_opsopsrngprimsdefault_tags__doc__ _prims_common RETURN_TYPENEW return_typer impl_atenprim_meta_impl) namerr* impl_metadoctags rngprim_def prim_packetprimps rregister_rng_primr4s--))tYR*Ki(%**..1148K   D    ++77;; = $ !rshapecp[R"[R"S[RS95$)Nrdtype)r TensorLikertensorint64)r5s rphilox_rand_offset_metar<3s$   U\\!5;;? @@rcSnUHnX-nM [R"U[RS9nSnSnSn[RR [RR 55nUR U-n[[U5n X-S- U-n [XRU-5n U S- XJ-U--S-U-$)Nr7) r scalar_tensorr;cudaget_device_propertiescurrent_devicemax_threads_per_multi_processorrintminmulti_processor_count) r5 numel_scalardim_sizenumel block_sizeunrollcurand4_engine_callsdevice_property blocks_per_smnum grid_sizes rphilox_rand_offsetrS9s L     EKK @EJ Fjj66uzz7P7P7RSO#CCzQM sE C!A%*4IIDD}TUI 1W*069 :Q >BV VVrc SnSnS[RS[RS[RS[[S4S-S[ S [ 4 S jnS[RS[RS[RS[[S4S-S[ S [ 4 S jn[UUUUS [RR4S 9 g)N philox_randz{(SymInt[] size, Tensor seed, Tensor offset, int[]? stride, Device? device=None, ScalarType? dtype=None) -> (Tensor, Tensor)r5seedoffsetstride.rr8cUb[SU35e[U5n[R"XXTS9n[ U5nXb4$)Nstride must be None, got )r5stridesr8r)AssertionErrorr r TensorMetar<)r5rVrWrXrr8 random_valuess r_philox_rand_meta/register_philox_rand.._philox_rand_metaSsQ   #._philox_randes   #get_device..s ,GS&=GrBc3*# UH oS:Hv M g7f)xpuNrrrs rrurv -WcE\Wrwryc3*# UH oS:Hv M g7f)hpuNrrrs rrurvrzrwr|c3*# UH oS:Hv M g7f)rbNrrrs rrurvrzrwrb)get isinstancestrrrrrkany)argskwargsrargrgs r get_devicers zz(H% fc " "\\&)F{{*.P$3*S%,,2Oszz$GP ,G ,,, -W - - - -W - - - -W - - - Qs ##D Dc^^^^^^"SS[5nU"5mTR[R5"[ TSS95 TR[R 5S5mTR[R 5S5mTR[R5S5mTR[R5S5mTR[R5UUUU4S j5mTR[5U4S j5nTR[5UU4S j5nT$) Nc4^\rSrSrU4SjrU4SjrSrU=r$)>register_run_and_save_rng_state_op..RunAndSaveRngStatec">[TU]SSS9 g)Nrun_and_save_rng_stateT cacheablesuper__init__self __class__s rrGregister_run_and_save_rng_state_op..RunAndSaveRngState.__init__s G 5  Frc,>[TU]"U/UQ70UD6$Nr__call__)roprrrs rrGregister_run_and_save_rng_state_op..RunAndSaveRngState.__call__s7#B888 8rr__name__ __module__ __qualname____firstlineno__rr__static_attributes__ __classcell__rs@rRunAndSaveRngStaters G 9 9rrTdeferred_errorcN[RR5U"U0UD64$r)rrB get_rng_staterrrs r impl_cuda5register_run_and_save_rng_state_op..impl_cudas$zz'')2t+>v+>>>rc<[R"5U"U0UD64$r)rrrs rimpl_cpu4register_run_and_save_rng_state_op..impl_cpus ""$b$&9&&999rc[[S5(a&[RR5U"U0UD64$[ S5eNr|z2functionalize a hpu RNG operator is not supported.)hasattrrr|rrrs rimpl_hpu4register_run_and_save_rng_state_op..impl_hpus= 5% 99**,b$.A&.AA AOPPrcN[RR5U"U0UD64$r)rryrrs rimpl_xpu4register_run_and_save_rng_state_op..impl_xpus$yy&&("d*=f*===rcn>TTTT S.n[X5nXC;a[SU35eX4nU"U/UQ70UD6$N)rBrbr|ryzBackend not supported for rr\) rrrimpl_maprimplrrrrs rimpl_backend_select?register_run_and_save_rng_state_op..impl_backend_selectsZ   D)  ! #=fX!FG GB((((rcV>U T"U/UQ70UD6sSSS5 $!,(df  g=frr)moderrrrs rimpl_fake_tensor_modeAregister_run_and_save_rng_state_op..impl_fake_tensor_modes$&r;D;F;TTs  (c4>T"U/UQ70UD6n[R"URRU/UQ75n[R"URRU5nURR ST XV5n[ XGSURS9$N call_functionconstanttracer)pytreetree_mapr unwrap_proxy create_proxyr ) rrrrout proxy_args proxy_kwargs out_proxyrrs rimpl_proxy_dispatch_modeDregister_run_and_save_rng_state_op..impl_proxy_dispatch_modes!"6t6v6__T[[%=%={T{K t{{'?'?H KK,, 3Z !$t{{SSr) rpy_implrAutogradrCUDACPUHPUXPU BackendSelectr r ) rrrrrrrrrs @@@@@@r"register_run_and_save_rng_state_oprs=90901"";#7#78 !7M##K$4$45?6?##KOO4:5:##KOO4Q5Q ##KOO4>5>##K$=$=> )? )##N3<4< ##$:;T<T "!rc^^^^^ "SS[5nU"5m T R[R5"[ T SS95 T R[R 5S5mT R[R 5S5mT R[R5S5mT R[R5S5mT R[5U 4S j5nT R[R5UUUU4S j5nT R[5S 5nT RU 4S j5nT $) Nc4^\rSrSrU4SjrU4SjrSrU=r$)7register_run_with_rng_state_op..RunWithRngStatec">[TU]SSS9 g)Nrun_with_rng_stateTrrrs rr@register_run_with_rng_state_op..RunWithRngState.__init__s G 1T  Brc,>[TU]"X/UQ70UD6$rr)r rng_staterrrrs rr@register_run_with_rng_state_op..RunWithRngState.__call__s7#ICDCFC Crrrrs@rRunWithRngStaters C D DrrTrc[RR5n[RRUR 55 U"U0UD6n[RRU5 U$r)rrBr set_rng_staterbrrrr current_staters rr1register_run_with_rng_state_op..impl_cudasR 002    1$!&!   / rc[R"5n[R"U5 U"U0UD6n[R"U5 U$r)rrrrs rr0register_run_with_rng_state_op..impl_cpus@++-  I&$!&! M* rc[[S5(af[RR5n[RR U5 U"U0UD6n[RR U5 U$[ S5er)rrr|rrrrs rr0register_run_with_rng_state_op..impl_hpuse 5% !II335M II # #I .d%f%C II # #M 2JOPPrc[RR5n[RRU5 U"U0UD6n[RRU5 U$r)rryrrrs rr0register_run_with_rng_state_op..impl_xpusL //1   *$!&!  . rc~>[5 T "X/UQ70UD6nSSS5 [R"URRX/UQ75n[R"URRU5nURR ST Xg5n[ WUSURS9$!,(df  N=frr rrrrrr ) rrrrrrrrrrs rr@register_run_with_rng_state_op..impl_proxy_dispatch_modes) *$YDTDVDC+__T[[%=%= ?UPT?UV t{{'?'?H KK,, / !i$t{{SS+ *s B.. B<cn>TTT T S.n[X#5nXT;a[SU35eXEnU"X/UQ70UD6$rr) rrrrrrrrrrrs rr;register_run_with_rng_state_op..impl_backend_selectsZ   D)  ! #=fX!FG GI3D3F33rcLU U"U0UD6sSSS5 $!,(df  g=frr)rrrrrs rr=register_run_with_rng_state_op..impl_fake_tensor_mode)st&v&TT #c>URU5nURU5nURU5nUR5 T "XR/UQ70UD6nURU5sSSS5 $!,(df  g=frunwrap_tensorsredispatch_to_next wrap_tensors) ctxrrrrunwrapped_rng_stateunwrapped_argsunwrapped_kwargsrrs rimpl_functional7register_run_with_rng_state_op..impl_functional0sz!00;++D1--f5  # # %$#*8.GraphSafeRunWithRngStateiDc$>[TU]S5 g)Ngraphsafe_run_with_rng_staterrs rrSregister_graphsafe_run_with_rng_state_op..GraphSafeRunWithRngState.__init__Es G ; [TU]"U/UQ7SU0UD6$Nrr)rrrrrrs rrSregister_graphsafe_run_with_rng_state_op..GraphSafeRunWithRngState.__call__Hs"7#BMMMfM Mrrrrs@rGraphSafeRunWithRngStaterDs =15 N NrrTrr cURRn[RRUnUR 5nUR U5 U"U0UD6nUR U5 U$r)rindexrrBdefault_generatorsgraphsafe_get_stategraphsafe_set_state)rrrr device_idx generatorrrs rr;register_graphsafe_run_with_rng_state_op..impl_cudaRsg%%++ JJ11*= !557 %%i0$!&!%%m4 rc^>[X#5nUS:wa[SU35eT"U/UQ7SU0UD6$)NrBz6GraphSafe RNG operations only supported for CUDA, got rr)rrrrrrs rrEregister_graphsafe_run_with_rng_state_op..impl_backend_select^sHD) V  HQ BdBiB6BBrcLU U"U0UD6sSSS5 $!,(df  g=frr)rrrrrs rrGregister_graphsafe_run_with_rng_state_op..impl_fake_tensor_modegs t&v&TTrc>[5 T "U/UQ7SU0UD6nSSS5 [R"URRU/UQ75n[R"URRSU0UE5nURR ST Xg5n[ WUSURS9$!,(df  N=f)Nrrrr) rrrrrrrrrrs rrJregister_graphsafe_run_with_rng_state_op..impl_proxy_dispatch_models ( *.rXDXIXQWXC+__T[[%=%={T{K  KK $ ${I&H&H KK,, 9: !i$t{{SS+ *s B44 Cc>UbURU5OSnURU5nURU5nUR5 T "U/UQ7SU0UD6nURU5sSSS5 $!,(df  g=fr r) rrrrrrrrrrs rrAregister_graphsafe_run_with_rng_state_op..impl_functionalys.7-BC  y ) ++D1--f5  # # %.#/BFVC##C( & % %s A33 B) rrrrrrrr r r)rrrrrrrs @@r(register_graphsafe_run_with_rng_state_oprCsN#6N$<#=  (()=)=> !=dS"))+*:*:;'+ < "))+*C*CD15CEC")).99=':'"))*@A<@ TB T"7726 )8 ) ('rc[5 gr)rorrrregister_rng_primsr!srr)(typingrrtorch.utils._pytreeutils_pytreerrtorch._Crtorch._higher_order_ops.utilsr torch._opsrtorch._prims_commonrr torch._subclasses.fake_tensorr "torch.fx.experimental.proxy_tensorr r r torch.typesrrrr4rjr<rSrorrrrrrrr!rrrr-s $$ B*O8 (%.A ::A W ::W,6r&?"D_D<=35D(N HIr