+ jK6 P^RIHt^RIt^RIt^RIt^RIHtHt^RIt^RI H t R]PP'dRMRRR R R R R1R R2/t ^@tR3t^tR ^R^R^R^/t^t^tR4RRlltRRlt!RR]4t!RR]4t] !4'dq^RIt^RIt^RIt]!]P8P:R4'd ^RIHtR5RltMN]!]P:P:R4'd ^RIHtR5RltMR5RltM]P@!R.R6O.R7OR7t!RR ]4t!!R!R"]4t"!R#R$]PF4t$R%R&lt%!R'R(]PLR)R*7t'!R+R,]PLR)R*7t(!R-R.]PF4t)!R/R0]PF4t*R#)8) annotationsN)autoEnum)has_triton_packageXi iYZR0_R1_XBLOCKYBLOCKZBLOCKR0_BLOCKc$V^8dQhRRRRRR/#)kwargsztyping.Mapping[str, int]r0_block int | Nonereturnint)formats"^/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/_inductor/runtime/hints.py __annotate__r#s$ $0:cVPR^4VPR^4,VPR^4,VfVPR^4,#T,#)r rrr)get)rrs&&rnative_matmul_block_numelr#si  8Q **Xq ! " **Xq ! ")1(86::j! $ H?G Hrc V^8dQhRRRR/#)rrrrr)rs"rrr.s//c/c/rc"\V\4#N)maxTRITON_DOT_MIN_BLOCK)rs&rnative_matmul_persistent_rblockr%.s x- ..rc&]tRt^2t^t^t^t^tRtR#) ReductionHintrN) __name__ __module__ __qualname____firstlineno__INNEROUTER OUTER_TINYDEFAULT__static_attributes__rrrr'r'2s E EJGrr'c]tRt^9t^t^tRtR#)TileHintrN)r(r)r*r+SQUAREr/r0rrrr2r29s FGrr2AttrsDescriptorr4c>RVRV/p\P!RVR\P/4pVPR,^8wd \ RVPR, 24hVPR,^8wd \ RVPR, 24hV#)tt.divisibilityz tt.equal_toarg_propertiesclsz$Expected tt.divisibility == 16, got zExpected tt.equal_to == 1, got )r4 from_dictr(property_valuesAssertionError)divisible_by_16 equal_to_1pointer_range_32rress&&& rAttrsDescriptorWrapperrAHs"?zF "++!65/2J2JKC""#45;$:3;N;NO`;a:bc""=1Q6$5c6I6I-6X5YZJrc$RVRV/p\R/VB#)r=r>rr5)r=r>r?rs&&& rrArAes%"?jF #,V, ,rcT;'gRUu/uF q33R^..bK ppT;'gRF.pV3pWT9dWE,PR^ .4K'R^ ..WE&K0 V#uupi)r7ztt.pointer_ranger)append)r=r>r?xresultkeys&&& rrArAysAP@U@USU@UW@U!d/455@UFW%+++d=K&&(:B'?@$6#;"> %>> ! 99 '$)),TUV VrcV^8dQhRR/#)rrrXr)rs"rrres   /  rc ^RIp^RIHpVPpVPP 'd VR8XdRpV!V4pVP V4pVPpT!TTPTTPT4\TRR4\TRR4\TR R4\TR R 4\TR R4R 7 # \d#TR8XdTPpLTR8Xd^@pLhi;i)rN)get_interface_for_devicergrhxpumtiar^r_r`rar rb) rZr[r\r]r^r_r`rarb) torchtorch._dynamo.device_interfacerprZversionrhget_device_propertiesr\AttributeErrorgpu_subslice_countr[get_compute_capabilitygetattr)r9devicersrp device_typedevice_interfacepropsr\s&& rcreateDeviceProperties.creates Kkk ==   !6K3F; 66v> $)$?$? !,,"766v>%$/$+E3Ld$S,38$-#*%1H$"Oe[$7   e#(-(@(@%&(*%  s CC0# C0.C0r)r(r)r*r+__doc____annotations__r^r_r`rarbpropertyrl classmethod functoolscacherr0rrrrXrXsY I J GE:*.Z.26#Z6(,:, Iz  __    rrXcV^8dQhRR/#rdr)rs"rrrs@@S@rc@\PV4P#)aUReturn the wave/warp size in threads for the given device. Reads from torch.cuda.get_device_properties(device).warp_size via the cached DeviceProperties.create(). Correct on both AMD (64 for CDNA/gfx9, 32 for RDNA/gfx10+) and NVIDIA (always 32). Missing cuda/hip warp_size metadata is treated as an error rather than silently falling back. )rXrrl)r{s&r get_warp_sizers  " "6 * ? ??rc]tRt^t$RtR]R&R]R&R]R&R]R&R ]R &R ]R &R ]R &R ]R&R ]R&R ]R&R]R&R]R&R]R&R]R&R]R&R]R&RtR#) TritonMetaa*Metadata bag threaded from Triton codegen into the runtime launcher. total=False because the key set is populated incrementally across codegen (signature/device/constants/configs first, then backend/ROCm/tlx extras) and the whole bag is forwarded verbatim to external Triton APIs, which tolerate and ignore keys they do not recognize. `device` is typed as the codegen-time DeviceProperties; CachingAutotuner rewrites it to the integer device index before reaching Triton, and the runtime read sites that expect that int narrow it with an explicit cast. dict[str, typing.Any] signaturerXr{rYr| constantszlist[typing.Any]configsbool native_matmullaunch_cooperative_gridenable_fp_fusion launch_pdl disable_ftzrmatrix_instr_nonkdim waves_per_eukpackztuple[str, ...] restore_value reset_to_zerobackend_optionsrNr(r)r*r+rrr0rrrrrsh %$ $$ !! J""""**rrF)totalc]tRtRt$RtR]R&R]R&R]R&R]R&R ]R &R ]R &R]R &R]R &R]R&R]R&R]R&R]R&R]R&R]R&R]R&R]R&R]R&R]R&R]R&R]R&R ]R&R]R&R]R&R]R&R]R &R]R!&R]R"&R]R#&R]R$&R]R%&R]R&&R]R'&R]R(&R]R)&R*]R+&R]R,&R]R-&R.]R/&R]R0&R]R1&R2]R3&R*]R4&R]R5&R]R6&R]R7&R]R8&R]R9&R]R:&R]R;&R<]R=&R]R>&R]R?&R]R@&R*]RA&R]RB&RC]RD&RC]RE&R*]RF&RG]RH&R*]RI&RJ]RK&R.]RL&R*]RM&R*]RN&R*]RO&R*]RP&R*]RQ&R]RR&RStRT#)U InductorMetai aThe inductor kernel-config / heuristics metadata bag. Produced on the codegen side (TritonKernel.inductor_meta_common / inductor_meta_per_kernel plus the codegen_kernel literal) and consumed by the runtime autotuning machinery in triton_heuristics.py, coordinate_descent_tuner.py, and autotune_cache.py. Every key is optional (total=False): consumers read via .get(...) with defaults, and several keys are injected only on specific paths (reductions, combo kernels, fixed grids). Dynamically keyed nested bags (e.g. combo_grid_meta) stay typed dict[str, Any] because their keys are computed at runtime. The codegen-side producers still build this bag as a plain dict[str, Any], so the write side is not yet checked against this TypedDict; typing the producers is left as a follow-up. str | None backend_hashrassert_indirect_indexingautotune_local_cacheautotune_pointwisez bool | Noneautotune_remote_cachebundled_autotune_remote_cacheforce_disable_cachesdynamic_scale_rblockincremental_autotune max_autotunemax_autotune_pointwisermin_split_scan_rblockspill_threshold store_cubin deterministicbatch_invariantforce_filter_reduction_configs&mix_order_reduction_allow_multi_stagesdynamic_disable_pipelining$are_deterministic_algorithms_enabledis_hip is_fbcodeprofile_bandwidthrYprofile_bandwidth_regexprofile_bandwidth_output/profile_bandwidth_with_do_bench_using_profilingcoordinate_descent_tuning coordinate_descent_search_radius'coordinate_descent_check_all_directionsno_x_dimatomic_add_foundnum_load num_store num_reductionz typing.Anyautotune_hints RSPLIT_SIZE"has_loadstore_with_contiguous_rdimzdict[str, int]tma_min_block_sizesuses_tmauses_device_tmaz dict[str, dict[str, typing.Any]]host_tma_descriptor_args tiling_scores min_xblock min_rblockpersistent_reductionr%add_persistent_rblockmax_persistent_rblockstrict_sum_rblockfloat kernel_num_gb kernel_flop grid_type kernel_namemutated_arg_names optimize_memrrbrareduction_hintrcombo_grid_metacombo_tuning_groups list[str]combo_coordesc_field_ordercombo_coordesc_field_limitscombo_warp_stage_candidatesextra_launcher_args fixed_gridprecomputed_grids config_argsuse_fast_triton_launcherrNrrrrrr s ""&&#..  $((,00 $$*.. O  !!599##&))-11NMN(,,''N>>OO%((N!!%%**## ))!//!++##!!""rrc]tRtRt$R]R&R]R&RtR]R&RtR]R&RtR ]R &RtR ]R &R R lt RRlt RRlt RRlt Rt R#)HalideInputSpecigrYctypenameNzlist[str] | Noneshapestrideroffsetalias_ofcV^8dQhRR/#rrrYr)rs"rrHalideInputSpec.__annotate__ossrc @VPR9dR#VP#) at::Half*z uint16_t*)r at::BFloat16*)rrjs&r bindings_typeHalideInputSpec.bindings_typeos ::7 7zzrcV^8dQhRR/#rr)rs"rrrtsBBSBrc VPR8XdR#VPR8XdR#RVPPRR4 R2#)rz$halide_type_t(halide_type_float, 16)rz%halide_type_t(halide_type_bfloat, 16)zhalide_type_of<*z>())rreplacerjs&r halide_typeHalideInputSpec.halide_typetsA :: $9 :: (: !3!3C!< =SAArcV^8dQhRR/#rrrr)rs"rrr{s""4"rc VPRJ#r"rrjs&r is_scalarHalideInputSpec.is_scalar{szzT!!rcV^8dQhRR/#rr)rs"rrr~s&&4&rc VPRJ#r"rrjs&r is_bufferHalideInputSpec.is_buffer~szz%%rr)r(r)r*r+rrrrrrrrrr0rrrrrgsP J I"E "#F #FJHj B"&&rrcn]tRtRt$R]R&R]R&RtR]R&RtR ]R &RtR ]R &R RltRRlt Rt R#) HalideMetaizlist[HalideInputSpec]argtypesrYtargetNr schedulerzdict[str, int | str] | Nonescheduler_flagsr cuda_devicecV^8dQhRR/#)rrrr)rs"rrHalideMeta.__annotate__s  i rc XRVP 2.pVP'dVPRVP 24VP'dXVP'g \ R4hVPP 4Fwr#VPRV RV 24K V#)z-Command line args to pass to halide generatorztarget=zautoscheduler=z,scheduler_flags requires scheduler to be setzautoscheduler.=)rr rDr r<items)rkargskvs& rrHalideMeta.argss$++'( >>> KK.(89 :    >>>$%STT,,224 nQCq455 rcV^8dQhRR/#rr)rs"rrr s,,,rc VPRJ#r")r rjs&ris_cudaHalideMeta.is_cudast++rr) r(r)r*r+rr r r rrr0rrrrrs;## K Iz 37O07"K" ,,rriiir")NNN)r=r>r?)rrr)+ __future__r collectionsrtypingenumrrrstorch.utils._tritonrrurhTRITON_MAX_BLOCKTRITON_MAX_RSPLITTRITON_MAX_TENSOR_NUMELr$TRITON_DEFAULT_BLOCK_SIZESTRITON_DEFAULT_RSPLITTRITON_DEFAULT_RSPLIT_SIZErr%r'r2tritontriton.backends.compilertriton.compiler.compilerhasattrbackendscompilerr4rA namedtuplerJrS NamedTuplerXr TypedDictrrrrrrrr-s"  2 """ 9 9 ! c a a  /Dt ##v''):;;< 2 ))+< = =< -( &)33= D47 v((7 t@+!!+