+ j ^RIt^RIHtHt^RIHtHt^RIHtH t ^RI H t ^RI H t ]PPt] !]P ]P"]P$]P&.4t] !.]P*N]P,N]P.N]P0N]P2N]P4N]P6N]P8N]P:N]P<N]P>N]P@N]PBN]PDN]PFN]PHN]PJN]PLN]PNN]PPN]PRN]PTN]PVN]PXN]PZN]P\N]P^N]P`N]PbN]PdN]PfN4t4] !]Pj]Pl]Pn]Pp]Pr]Pt]Pv]Px]Pz. 4t>]4]>,t?RRRllt@RR R lltAR R ltBRR RlltCR#)N)datasheet_dram_bw_gbsdatasheet_tops)get_device_tflopsget_gpu_dram_gbps)optimization_hintstatically_known_true) OrderedSet) flop_registrycdV^8dQhR\\,RRR\R,R\/#)flopsdtypez torch.dtypegpu_typeNreturn)floatintstr)formats"`/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/utils/_runtime_estimation.py __annotate__rNs422 3;2,28;d 2 2c.VeV\PPPPR8Hp\ WVR7pVf\ RV: RV 24hM \V4pVR,pV^8XdR#V^, pVRV,, R,#) a>Convert a FLOPs count to estimated nanoseconds on the GPU. Uses 75% of theoretical peak and converts FLOPs to MACs (divide by 2). If ``gpu_type`` names a device in the datasheet, its pinned datasheet TFLOPS are used instead of querying the current device, making the estimate deterministic and hardware-independent. tf32)is_tf32 device_name gpu_type z has no datasheet entry for gmBg?geA)torchbackendscudamatmulfp32_precisionr ValueErrorr)r rrrtflopspeak_gpu_flopsmacss&&& r flops_to_nsr'Ns..%%,,;;vEHM >H<'CE7K   #5)d]N 19D D>) *c 11rc$V^8dQhR\/#r rr)rs"rrris!! !rcV\9dd\V4^8wd\RV RV 24hVP4p\V,p/p VeWYR&V!V/VBRV/BV Bp \ WVR7#R#)at Estimates the compute time of an aten operator. Args: func_packet: The operator overload packet. args: The arguments to the operator. kwargs: The keyword arguments to the operator. out: The output of the operator. out_dtypes: The output data types. node_meta: Optional FX node meta dict. Passed through to the flop formula as ``_node_meta`` kwarg so formulas can read annotations like ``sparsity_hint``. gpu_type: Optional datasheet device name to pin the peak FLOPS to instead of querying the current device. Returns: float: The estimated compute time in nanoseconds. z"Only support single out dtype got z for _node_metaout_val)rr)r lenAssertionErrorpopr') func_packetargskwargsout out_dtypes node_metarrflop_count_func extra_kwargs flop_counts &&&&&&& rget_compute_timer:is*m# z?a  4ZLk]S  ' 4  )2 &$dRfRcR\R :x@@ rcDV^8dQhR\PR\/#)r tr)rTensorr)rs"rrrs))U\\)c)rc^p\VPVP44F/wr#\V^8H4'dKV\ V^R7,pK1 WP 4,#)z Calculates the memory consumption of a tensor. Args: t (torch.Tensor): The input tensor. Returns: int: The memory consumption of the tensor in bytes. )fallback)zipshapestriderr element_size)r< real_numelsizerBs& r get_num_bytesrFsYJAGGQXXZ0 $Vq[11 +D1= =J1 ( ((rc$V^8dQhR\/#r)r*)rs"rrrsUrcVe!\V4pVf\RV: R24hM \4p\RV44p\RV44pWE,pWc, pV#)a Estimates the memory transfer time of input and output tensors. Args: flat_args_kwargs (List[torch.Tensor]): The flat list of arguments and keyword arguments. flat_outs (List[torch.Tensor]): The flat list of outputs. gpu_type: Optional datasheet device name to pin the DRAM bandwidth to instead of querying the current device. Returns: float: The estimated memory transfer time in nanoseconds. rz not found in datasheetc3|"TF2p\V\P4'gK%\V4xK4 R#5iN isinstancerr=rF.0r<s& r $get_transfer_time..s,"2QjELL6Q a"2"<<c3|"TF2p\V\P4'gK%\V4xK4 R#5irJrKrMs& rrOrPs+"+Qz!U\\/J a)rQ)rr#rsum)flat_args_kwargs flat_outsrgpu_memory_bandwidth read_bytes write_bytes counted_bytes transfer_times&&& rget_transfer_timer[s4X>  'y 4KLM M ( 12"2J"+K,M!8M rrJ)NN)Dr$torch._inductor.analysis.device_inforrtorch._inductor.utilsrr%torch.fx.experimental.symbolic_shapesrrtorch.utils._ordered_setr flop_counterr opsatenfloat16bfloat16float32float64 _FLOAT_TYPES lift_freshr< transposeviewdetach _unsafe_viewsplitadjoint as_strideddiagonalexpand expand_asmovedimpermuteselectsqueezemTmHrealimagview_as unflattenunfoldunbind unsqueezevsplithsplitsplit_with_sizesswapaxesswapdimschunk _VIEW_OPSrandintrandnrand randn_like rand_like randint_likearange ones_like zeros_like _CREATE_OPS _IGNORE_OPSr'r:rFr[rrrs VF0' yy~~                                    ! " # $ % & ' ( ) * + , - . / 0 1 2 3 4 5 6 7 8 9 : ; < = > ? "  H            +% 26!H)&r