Skip to main content

Getting StartedPlayground Contact

IntroductionType SafetyTensor IndexingEinsumEinopsAutogradTraining a ModelProfiling & MemoryPyTorch MigrationBest PracticesRuntimesPerformance

allow_mutation_on_saved_tensorsapplybackwardBackwardOptionsdetect_anomalydisable_saved_tensors_hooksemit_nvtxEmitNvtxContextenable_gradexecuteBackwardexport_chrome_traceFunctionFunctionCtxgradgrad_modegradcheckGradcheckOptionsGradcheckResultgradgradcheckGradOptionshessianhvpinference_modeis_anomaly_detection_enabledis_grad_enabledis_inference_mode_enabledjacobianjvpkey_averagesmark_dirtymark_non_differentiableno_gradNodeonce_differentiablePackHookprofileProfilerContextProfilerKeyAveragesProfilerOptionssave_for_backwardsave_on_cpusaved_tensors_hooksset_detect_anomalyset_grad_enabledset_materialize_gradsset_multithreading_enabledstartstoptableUnpackHookvhpvjp

absacosacoshaddaddbmmAddbmmOptionsaddcdivAddcdivOptionsaddcmulAddcmulOptionsaddmmAddmmOptionsaddmvAddmvOptionsaddrAddrOptionsadjointallallcloseAllcloseOptionsamaxaminaminmaxangleanyapplyOutarangeare_deterministic_algorithms_enabledargmaxargminargsortargwhereas_stridedas_tensorasinasinhAssertNoShapeErrorAtat_error_index_out_of_boundsatanatan2atanhatleast_1datleast_2datleast_3dAtShapeautocast_decrement_nestingautocast_increment_nestingAxesRecordbaddbmmBaddbmmOptionsbatch_dimensions_do_not_match_errorbernoulliBinaryOptionsbincountbitwise_andbitwise_left_shiftbitwise_notbitwise_orbitwise_right_shiftbitwise_xorblock_diagbmmbroadcast_error_incompatible_dimensionsbroadcast_shapesbroadcast_tensorsbroadcast_toBroadcastShapebroadcastShapesbucketizecanBroadcastTocartesian_prodcatCatShapecdistceilchain_matmulCholeskyShapechunkchunk_error_dim_out_of_rangeclampclear_autocast_cacheclonecolumn_stackcombinationscompiled_with_cxx11_abicomplexconjconj_physicalcontiguouscopysigncorrcoefcoscoshcount_nonzerocovCPUTensorDatacreateTorchCumExtremeResultcummaxcummincumprodCumShapecumsumcumulative_trapezoidCumulativeOptionsdeg2raddetachDetShapeDeviceDeviceInputDeviceTypediagdiag_embeddiagflatdiagonal_scatterDiagShapediffdigammadimension_error_out_of_rangedistdivdotdsplitdstackDTypeDynamicShapeEigShapeeinops_error_ambiguous_decompositioneinops_error_anonymous_in_outputeinops_error_dimension_mismatcheinops_error_invalid_patterneinops_error_reduce_undefined_outputeinops_error_repeat_missing_sizeeinops_error_undefined_axiseinsumeinsum_error_dimension_mismatcheinsum_error_index_out_of_rangeeinsum_error_invalid_equationeinsum_error_invalid_sublist_elementeinsum_error_operand_count_mismatcheinsum_error_subscript_rank_mismatcheinsum_error_unknown_output_indexEinsumOutputShapeEllipsiseluembedding_bag_error_requires_2d_inputemptyempty_cacheempty_likeeqequalerferfcerfinvexpexp2expandexpand_asexpand_error_incompatibleExpandShapeexpm1eyeEyeOptionsflattenFlattenShapeflipflip_error_dim_out_of_rangefliplrFlipShapeflipudfloat_powerfloorfloor_dividefmaxfminfmodfracfrexpfrombufferfullfull_likegathergather_error_dim_out_of_rangeGatherShapegcdgegeluget_autocast_cpu_dtypeget_autocast_gpu_dtypeget_autocast_ipu_dtypeget_autocast_xla_dtypeget_default_deviceget_default_dtypeget_deterministic_debug_modeget_device_moduleget_file_pathget_float32_matmul_precisionget_num_interop_threadsget_num_threadsget_printoptionsget_rng_stateGradFngthardsigmoidhardswishHasShapeErrorheavisidehistchistogramHistogramResulthsplithstackhypoti0imagindex_addindex_copyindex_fillindex_putindex_reduceindex_selectindex_select_error_dim_out_of_rangeIndexSelectShapeIndexSpecIndicesSpecinverseInverseShapeis_anomaly_check_nan_enabledis_anomaly_enabledis_autocast_cache_enabledis_autocast_cpu_enabledis_autocast_ipu_enabledis_autocast_xla_enabledis_complexis_complex_dtypeis_cpu_only_modeis_deterministic_algorithms_warn_only_enabledis_floating_pointis_floating_point_dtypeis_inference_mode_enabledis_nonzerois_tensoris_warn_always_enabledis_webgpu_availableIs2DIsAtLeast1DiscloseIscloseOptionsisfiniteisinisinfisnanisneginfisposinfisrealIsShapeErroritem_error_not_scalarItemResultkronkthvalueKthvalueOptionslcmldexpleleaky_relulerplgammalinalg_error_not_square_matrixlinalg_error_requires_2dlinalg_error_requires_at_least_2dlinspaceloglog10log1plog2logaddexplogaddexp2logcumsumexplogical_andlogical_notlogical_orlogical_xorlogitlogspacelogsumexpltLUShapemasked_selectmasked_select_asyncMaskSpecmatmulmatmul_error_inner_dimensions_do_not_matchMatmul2DShapeMatmulShapemaxmaximummeanmedianmemory_statsmemory_summarymeshgridminminimummmmodemovedimmsortmulmultinomialmultinomial_asyncmvnan_to_numnanmeannanmediannanquantilenansumnarrownarrow_copynarrow_error_length_exceeds_boundsnarrow_error_start_out_of_boundsNarrowShapeneneedsBroadcastnegNegativeDimnextafternonzeronormnormalNormOptionsnumelonesones_likeouterpackPackShapepermutepermute_error_dimension_count_mismatchPermuteShapepoissonpolarpositivepowPrintOptionsprodprofiler_allow_cudagraph_cupti_lazy_reinit_cuda12promote_typesquantileQuantileOptionsrad2degrandrand_likerandintrandint_likerandnrandn_likerandpermRangeSpecRankravelrealRearrangeShapereciprocalreduceReduceOperationReduceShapeReductionOptionsreluremainderrepeatrepeat_interleaveRepeatInterleaveOptionsRepeatShaperequireWebGPUreset_peak_memory_statsreshapeReshapeShaperesult_typerollrot90roundrsqrtscatterscatter_addscatter_add_scatter_error_dim_out_of_rangescatter_reducescatter_reduce_ScatterShapesearchsortedselectselect_error_index_out_of_boundsselect_scatterSelectShapeseluset_default_deviceset_default_tensor_typeset_deterministic_debug_modeset_float32_matmul_precisionset_printoptionsset_warn_alwaysShapeShapedTensorsigmoidsignsignbitsilusinsincsinhslice_error_out_of_boundsslice_scatterSliceShapeSliceSpecsoftmax_error_dim_out_of_rangeSoftmaxShapesoftplussoftsignsortSortOptionssplitsplit_error_dim_out_of_rangesqrtsquaresqueezeSqueezeShapestackstdstd_meanStdVarOptionssubSublistSublistElementSubscriptIndexsumSVDShapeswapaxessym_floatsym_intsym_notttaketake_along_dimtantanhtensortensor_splitTensorCreatorTensorDatatensordotTensorOptionsTensorStoragetileTileShapetopkTopkOptionsTorchtraceTraceShapetransposetranspose_dims_error_out_of_rangetranspose_error_requires_2d_tensorTransposeDimsShapeTransposeDimsShapeCheckedTransposeShapetrapezoidtriltril_indicestriutriu_indicestruncTypedArrayTypedStorageUnaryOptionsunbindunbind_error_dim_out_of_rangeunflattenuniqueunique_consecutiveunpackUnpackShapeunravel_indexunsqueezeUnsqueezeShapeuse_deterministic_algorithmsValidateBatchedSquareMatrixValidateChunkDimValidatedEinsumShapevalidateDeviceValidatedRearrangeShapeValidatedReduceShapeValidatedRepeatShapevalidateDTypeValidateEinsumValidateOperandCountValidateRanksValidateScalarValidateSplitDimValidateSquareMatrixValidateUnbindDimvar_var_meanvdotviewview_as_complexview_as_realvmapvsplitvstackWebGPUTensorDatawherexlogyzeroszeros_like

torch.js· 2026

Legal Terms of Use Privacy Policy

/

/

docs
torch.js
torch
autograd
executeBackward

torch.autograd.executeBackward

function executeBackward(root: Tensor, grad: Tensor, retainGraph: boolean): void

Execute the backward pass using proper topological ordering.

This is the core autograd engine that ensures gradients are properly accumulated before propagation. It implements the same algorithm as PyTorch:

Build the graph and compute dependency counts
Use a ready queue to process nodes in topological order
Accumulate gradients in input buffers until all dependencies satisfied

Parameters

rootTensor: The tensor to start backward from
gradTensor: The gradient of the loss w.r.t. root
retainGraphboolean: If true, keep the computation graph for another backward