Skip to main content
torch.js has not been released yet.
torch.js logotorch.js logotorch.js
PlaygroundContact
Login
Documentation
IntroductionType SafetyTensor ExpressionsTensor IndexingEinsumEinopsAutogradTraining a ModelProfiling & MemoryPyTorch MigrationBest PracticesRuntimesPerformancePyTorch CompatibilityBenchmarksDType Coverage
ActivationOptionsAdaptiveAvgPool1dAdaptiveAvgPool2dAdaptiveAvgPool3dAdaptiveLogSoftmaxOptionsAdaptiveLogSoftmaxWithLossAdaptiveMaxPool1dAdaptiveMaxPool1dOptionsAdaptiveMaxPool2dAdaptiveMaxPool2dOptionsAdaptiveMaxPool3dAdaptiveMaxPool3dOptionsadd_moduleAlphaDropoutappendappendapplyAvgPool1dAvgPool1dOptionsAvgPool2dAvgPool2dOptionsAvgPool3dAvgPool3dOptionsBackwardHookBackwardPreHookBatchNorm1dBatchNorm2dBatchNorm3dBatchNormOptionsBCELossBCEWithLogitsLossBilinearBilinearOptionsBufferBufferOptionsBufferRegistrationHookbufferscallCELUCELUOptionsChannelShufflechildrenCircularPad1dCircularPad2dCircularPad3dclearConstantPad1dConstantPad2dConstantPad3dConv1dConv2dConv3dConvOptionsConvTranspose1dConvTranspose2dConvTranspose3dConvTransposeOptionsCosineEmbeddingLossCosineEmbeddingLossOptionsCosineSimilarityCosineSimilarityOptionscreatecreateCrossEntropyLossCTCLossdecodedecodedeleteDropoutDropout1dDropout2dDropout3dDropoutOptionsELUELUOptionsEmbeddingEmbeddingBagEmbeddingBagForwardOptionsEmbeddingBagFromPretrainedOptionsEmbeddingBagOptionsEmbeddingFromPretrainedOptionsEmbeddingOptionsencodeencodeentriesentriesevalextendFeatureAlphaDropoutFlattenFlattenOptionsFoldFoldOptionsforwardforwardforwardforwardforwardforwardforwardforwardforwardforwardforwardforwardforward_with_targetForwardHookForwardPreHookFractionalMaxPool2dFractionalMaxPool3dFractionalMaxPoolOptionsfrom_pretrainedfrom_pretrainedGaussianNLLLossGELUGELUOptionsgenerate_square_subsequent_maskgetgetgetgetgetget_bufferget_parameterget_submoduleGLUGLUOptionsGroupNormGroupNormOptionsGRUGRUCellHardshrinkHardshrinkOptionsHardsigmoidHardswishHardtanhHardtanhOptionshashasHingeEmbeddingLossHingeEmbeddingLossOptionsHuberLossHuberLossOptionsIdentityInstanceNorm1dInstanceNorm2dInstanceNorm3dInstanceNormOptionsis_uninitialized_bufferis_uninitialized_parameteriterator]iterator]iterator]iterator]keyskeysKLDivLossL1LossL1LossOptionsLayerNormLayerNormOptionsLazyBatchNorm1dLazyBatchNorm2dLazyBatchNorm3dLazyConv1dLazyConv2dLazyConv3dLazyConvOptionsLazyConvTranspose1dLazyConvTranspose2dLazyConvTranspose3dLazyConvTransposeOptionsLazyInstanceNorm1dLazyInstanceNorm2dLazyInstanceNorm3dLazyLinearLeakyReLULeakyReLUOptionsLinearLinearOptionsload_state_dictload_state_dictLocalResponseNormLocalResponseNormOptionslog_probLogSigmoidLogSoftmaxLogSoftmaxOptionsLPPool1dLPPool1dOptionsLPPool2dLPPool2dOptionsLPPool3dLPPool3dOptionsLSTMLSTMCellLSTMCellOptionsMarginRankingLossMarginRankingLossOptionsmaterializematerializematerialize_uninitializedmaterialize_uninitializedMaxPool1dMaxPool1dOptionsMaxPool2dMaxPool2dOptionsMaxPool3dMaxPool3dOptionsMaxUnpool1dMaxUnpool1dOptionsMaxUnpool2dMaxUnpool2dOptionsMaxUnpool3dMaxUnpool3dOptionsMishModuleModuleBuffersModuleChildrenModuleDictModuleDictOptionsModuleListModuleListOptionsModuleParametersModuleRegistrationHookmodulesMSELossMSELossOptionsmultihead_attnMultiheadAttentionMultiheadAttentionOptionsMultiheadAttnOptionsMultiLabelMarginLossMultiLabelMarginLossOptionsMultiLabelSoftMarginLossMultiMarginLossnamed_buffersnamed_childrennamed_modulesnamed_parametersNamedModulesOptionsNamedRecurseOptionsNLLLossnum_parametersNumParametersOptionsPairwiseDistancePairwiseDistanceOptionsParameterParameterDictParameterDictOptionsParameterListParameterListOptionsParameterOptionsParameterRegistrationHookparametersPixelShufflePixelUnshufflePoissonNLLLosspoppopPReLUPReLUOptionsRecurseOptionsReflectionPad1dReflectionPad2dReflectionPad3dregister_backward_hookregister_bufferregister_forward_hookregister_forward_pre_hookregister_full_backward_hookregister_full_backward_pre_hookregister_module_backward_hookregister_module_buffer_registration_hookregister_module_forward_hookregister_module_forward_pre_hookregister_module_full_backward_hookregister_module_full_backward_pre_hookregister_module_module_registration_hookregister_module_parameter_registration_hookregister_parameterReLUReLU6RemovableHandleremoveReplicationPad1dReplicationPad2dReplicationPad3dRMSNormRMSNormOptionsRNNRNNBaseRNNBaseOptionsRNNCellRNNCellOptionsRReLURReLUOptionsrunrunSELUSequentialsetsetsetSigmoidSiLUSmoothL1LossSmoothL1LossOptionsSoftMarginLossSoftMarginLossOptionsSoftmaxSoftmax2dSoftmaxOptionsSoftminSoftminOptionsSoftplusSoftplusOptionsSoftshrinkSoftshrinkOptionsSoftsignstate_dictstate_dictStateDictOptionsstepSyncBatchNormTanhTanhshrinkThresholdThresholdOptionstotototrainTrainOptionsTransformerTransformerDecoderTransformerDecoderDecodeOptionsTransformerDecoderLayerTransformerDecoderLayerDecodeOptionsTransformerDecoderLayerOptionsTransformerDecoderOptionsTransformerEncoderTransformerEncoderEncodeOptionsTransformerEncoderLayerTransformerEncoderLayerEncodeOptionsTransformerEncoderLayerOptionsTransformerEncoderOptionsTransformerOptionsTransformerRunOptionsTripletMarginLossTripletMarginWithDistanceLossUnflattenUnfoldUnfoldOptionsUninitializedBufferUninitializedOptionsUninitializedParameterupdateUpsampleUpsamplingBilinear2dUpsamplingNearest2dvaluesvalueszero_gradZeroPad1dZeroPad2dZeroPad3d
absacosacoshAdaptivePool1dShapeAdaptivePool2dShapeaddaddbmmAddbmmOptionsaddcdivAddcdivOptionsaddcmulAddcmulOptionsaddmmAddmmOptionsaddmvAddmvOptionsaddrAddrOptionsadjointallallcloseAllcloseOptionsAlphaBetaOptionsamaxaminaminmaxAminmaxOptionsangleanyapplyOutarangeare_deterministic_algorithms_enabledargmaxargminargsortargwhereas_stridedas_tensorasinasinhAssertNoShapeErrorAssertNotErrorAsStridedOptionsAtat_error_index_out_of_boundsatanatan2atanhatleast_1datleast_2datleast_3dAtShapeautocast_decrement_nestingautocast_increment_nestingautograd_gradient_mismatch_errorautograd_not_registered_errorAutogradConfigAutogradDeviceAutogradDTypeAutogradEntryAutogradHandleAutogradHandleImplAxesRecordBackwardFnbaddbmmBaddbmmOptionsbartlett_windowBaseKernelConfigbatch_dimensions_do_not_match_errorbernoulliBernoulliOptionsBinaryBackwardFnBinaryBroadcastResultBinaryDTypeBinaryKernelConfigCPUBinaryKernelCPUBinaryOpConfigBinaryOpNamesBinaryOpSchemaBinaryOptionsbincountBincountOptionsbitwise_andbitwise_left_shiftbitwise_notbitwise_orbitwise_right_shiftbitwise_xorblackman_windowblock_diagbmmBooleanDTypeRulebroadcast_error_incompatible_dimensionsbroadcast_shapesbroadcast_tensorsbroadcast_toBroadcastShapeBroadcastShapeRulebroadcastShapesbucketizeBucketizeOptionsBufferUsagebuildEinopsErrorbuildErrorMessagecanBroadcastTocartesian_prodcatCatOptionsCatShapeCauchyOptionscdistCdistOptionsceilceluCeluFunctionalOptionschain_matmulCheckShapeErrorCholeskyShapechunkchunk_error_dim_out_of_rangeChunkOptionsclampClampOptionsclear_autocast_cacheclearEinopsCacheclearEinsumCacheclonecolumn_stackcombinationsCombinationsOptionscompiled_with_cxx11_abicomplexconjconj_physicalcontiguousConv1dShapeConv2dShapeConv3dShapeConvTranspose2dShapecopysigncorrcoefcoscoshcount_nonzeroCountNonzeroOptionscovcoverage_reportcoverageReportCoverageReportCovOptionsCPUForwardFnCPUKernelConfigCPUKernelEntryCPUOnlyResultCPUTensorDatacreateCumExtremeResultcreateTorchCreationOpSchemaCumExtremeResultcummaxcummincumprodCumShapecumsumcumulative_trapezoidCumulativeOptionsCumulativeOptionsWithDimdeg2raddetachDeterministicOptionsDetShapeDevicedevice_error_requiresDeviceBufferDeviceCapabilitiesDeviceCheckedResultDeviceConfigDeviceContextDeviceEntryDeviceHandleDeviceInputDeviceOptionsDeviceRegistryDeviceTypediagdiag_embedDiagEmbedOptionsdiagflatDiagflatOptionsDiagFlatOptionsdiagonal_scatterDiagonalOptionsDiagonalScatterOptionsDiagOptionsDiagShapediffDiffOptionsdigammadimension_error_out_of_rangeDispatchConfigdistDistOptionsdivdotDotShapeRuleDoubleDoubleDimdropoutDropoutFunctionalOptionsdsplitdstackDTypedtype_already_registered_errordtype_components_mismatch_errordtype_not_found_errorDTypeComponentsDTypeConfigDTypeCoverageReportDTypeDisplayConfigDTypeEntryDTypeHandleDTypeHandleImplDTypeInfoDTypeRegistryDTypeRuleDTypeSerializationConfigDynamicShapeEigShapeeinops_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_indexEinsumOptionsEinsumOutputShapeEllipsiseluelu_EluFunctionalOptionsembedding_bag_error_requires_2d_inputemptyempty_cacheempty_likeeqequalerferfcerfinvexpexp2expandexpand_asexpand_error_incompatibleExpandShapeexpm1ExponentialOptionseyeEyeOptionsfftFFTOptionsfindKernelWithPredicatefindSimilarPatternsflattenFlattenOptionsFlattenShapeflipflip_error_dim_out_of_rangefliplrFlipShapeflipudfloat_powerFloatDTypeRulefloorfloor_dividefmaxfminfmodformatEquationErrorformatShapefracfrexpfrombufferfullfull_likefunction_already_registered_errorFunctionConfigFunctionEntryFunctionHandlegathergather_error_dim_out_of_rangeGatherShapegcdgegeluGeometricOptionsget_autocast_cpu_dtypeget_autocast_gpu_dtypeget_autocast_ipu_dtypeget_autocast_xla_dtypeget_default_deviceget_default_dtypeget_deterministic_debug_modeget_device_configget_device_contextget_device_moduleget_dtype_infoget_file_pathget_float32_matmul_precisionget_num_interop_threadsget_num_threadsget_op_infoget_printoptionsget_real_dtypeget_rng_stategetAutogradgetDTypegetEinopsCacheSizegetEinsumCacheSizegetFunctiongetKernelgetMethodgetOpInfoGetOpKindGetOpSchemagetScalarKernelgluGluFunctionalOptionsGradContextGradFnGradientsForgtHalfHalfDimhamming_windowhann_windowhardshrinkhardsigmoidhardswishhardtanhhardtanh_HardtanhFunctionalOptionshas_autogradhas_devicehas_dtypehas_kernelhasAutogradhasDTypehasFunctionhasKernelhasMethodhasScalarKernelHasShapeErrorheavisidehistcHistcOptionshistogramHistogramOptionsHistogramResulthsplithstackhypoti0IdentityShapeifftimagindex_addindex_copyindex_fillindex_putindex_reduceindex_selectindex_select_error_dim_out_of_rangeIndexPutOptionsIndexSelectShapeIndexSpecIndicesOptionsIndicesSpecinitialize_deviceInputsForInsertDiminvalid_config_errorinverseInverseShapeirfftis_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_availableIs2DIsAtLeast1DIsBinaryOpIsBinaryOpNameiscloseIscloseOptionsisfiniteisinisinfisnanisneginfisposinfisrealIsReductionOpIsReductionOpNameIsRegistryErrorIsShapeErroristftISTFTOptionsIsUnaryOpIsUnaryOpNameitem_error_not_scalarItemResultkaiser_windowKaiserWindowOptionskernel_not_registered_errorkernel_signature_mismatch_errorKernelConfigKernelConfigWebGPUKernelEntryKernelHandleKernelInfoKernelPredicateKernelRegistryKernelWebGPUkronkthvalueKthvalueOptionslcmldexpleleaky_reluleaky_relu_LeakyReluFunctionalOptionslerplevenshteinDistancelgammalinalg_error_not_square_matrixlinalg_error_requires_2dlinalg_error_requires_at_least_2dlinearlinspacelist_custom_deviceslist_custom_dtypeslist_deviceslist_dtypeslist_functionslist_kernelslist_methodslist_opslistCustomDTypeslistDTypeslistFunctionslistKernelsListKernelsOptionslistMethodslistOpsListOpsOptionsloglog_softmaxlog10log1plog2logaddexplogaddexp2logcumsumexplogical_andlogical_notlogical_orlogical_xorLogitOptionsLogNormalOptionsLogOptionslogsigmoidlogspacelogsumexpLogsumexpOptionsltLUShapeLuSolveOptionsmasked_fillmasked_selectmasked_select_asyncMaskSpecmatmulmatmul_error_inner_dimensions_do_not_matchMatmul2DShapeMatmulShapeMatmulShapeRuleMatrixTransposeShapemaxmaximummeanmedianmemory_statsmemory_summarymeshgridmethod_already_registered_errormethod_dtype_not_supported_errorMethodConfigMethodEntryMethodHandleminminimummishmmMMShapeRulemodemovedimmsortmulmultinomialmultinomial_asyncMultinomialAsyncOptionsMultinomialOptionsMultiplyBymvMVShapeRulenan_to_numnanmeannanmediannanquantileNanReductionOptionsnansumNanToNumOptionsnarrownarrow_copynarrow_error_length_exceeds_boundsnarrow_error_start_out_of_boundsNarrowShapeneneedsBroadcastnegNegativeDimnextafternonzeroNonzeroOptionsnormnormalNormalOptionsNormOptionsnumelonesones_likeop_kind_mismatch_errorop_not_found_errorOpCoverageEntryOpInfoOpKindOpNameOpSchemaOpSchemasouterOuterShapepackPackShapepermutepermute_error_dimension_count_mismatchPermuteShapepoissonpolarPool1dShapePool2dShapePool3dShapepositivepowpreluPrintOptionsprodprofiler_allow_cudagraph_cupti_lazy_reinit_cuda12promote_typesPromoteDTypeRulePutOptionsquantileQuantileOptionsrad2degrandrand_likerandintrandint_likeRandintLikeOptionsRandintOptionsrandnrandn_likeRandomLikeOptionsRandomOptionsrandpermRangeSpecRankravelrealrearrangeRearrangeOptionsRearrangeShapereciprocalreduceReduceOperationReduceOptionsReduceShapeReductionKernelConfigCPUReductionKernelCPUReductionOpNamesReductionOpSchemaReductionOptionsReductionShapeRuleregister_backwardregister_deviceregister_dtyperegister_forwardregister_functionregister_methodregister_scalar_forwardregisterAutogradRegisterBackwardOptionsregisterBinaryOpregisterDTypeRegisterDTypeOptionsRegisteredDTyperegisterFunctionRegisterFunctionOptionsregisterKernelRegisterKernelOptionsregisterMethodRegisterMethodOptionsregisterScalarKernelregisterUnaryOpregistration_failed_errorrelurelu_relu6ReluFunctionalOptionsremainderRemoveDimrepeatrepeat_interleaveRepeatInterleaveOptionsRepeatOptionsRepeatShapeReplaceDimrequireWebGPUreset_peak_memory_statsreshapeReshapeShaperesult_typerfftrollRollOptionsrot90Rot90Optionsroundrrelurrelu_RreluFunctionalOptionsrsqrtSafeExpandShapeSameDTypeRuleSameShapeRuleSaveForBackwardScalarCPUForwardFnScalarCPUKernelConfigScalarKernelEntryScalarKernelHandleScalarWebGPUKernelConfigScaleDimscatterscatter_addscatter_add_scatter_error_dim_out_of_rangescatter_reducescatter_reduce_ScatterReduceOptionsScatterShapesearchsortedSearchSortedOptionsselectselect_error_index_out_of_boundsselect_scatterSelectShapeseluset_default_deviceset_default_tensor_typeset_deterministic_debug_modeset_float32_matmul_precisionset_printoptionsset_warn_alwaysSetupContextFnShapeShapeCheckedResultShapedTensorShapeErrorMessageShapeOpSchemaShapeRulesigmoidsignsignbitsilusinsincsinhSizeOptionsslice_error_out_of_boundsslice_scatterSliceOptionsSliceScatterOptionsSliceShapeSliceSpecsoftmaxsoftmax_error_dim_out_of_rangeSoftmaxShapesoftminSoftminFunctionalOptionssoftplusSoftplusFunctionalOptionssoftshrinksoftsignsortSortOptionssplitsplit_error_dim_out_of_rangeSplitOptionssqrtsquaresqueezeSqueezeOptionsSqueezeShapestackStackOptionsStackShapestdstd_meanStdVarMeanOptionsStdVarOptionsstftSTFTOptionsStrideOptionssubSublistSublistElementSubscriptIndexsumSVDShapeswapaxessym_floatsym_intsym_notttaketake_along_dimTakeAlongDimOptionstantanhtanhshrinktensortensor_splitTensorCreatorTensorDatatensordotTensordotOptionsTensorLikeTensorMetaTensorOptionsTensorStoragethresholdthreshold_tileTileShapeToOptionstopkTopkOptionsTorchtraceTraceShapetransposetranspose_dims_error_out_of_rangetranspose_error_requires_2d_tensorTransposeDimsShapeTransposeDimsShapeCheckedTransposeShapetrapezoidTrapezoidOptionsTriangularOptionstriltril_indicesTriOptionsTripletriutriu_indicestrue_dividetruncTupleOfLengthTypedArrayTypedArrayForTypedStorageTypeOptionsUnaryBackwardFnUnaryDTypeUnaryKernelConfigCPUUnaryKernelCPUUnaryOpConfigUnaryOpFnUnaryOpNamesUnaryOpParamsUnaryOpSchemaUnaryOptionsunbindunbind_error_dim_out_of_rangeUnbindOptionsunflattenUniformOptionsuniqueunique_consecutiveUniqueConsecutiveOptionsUniqueOptionsunpackUnpackShapeunravel_indexunregister_deviceunsqueezeUnsqueezeOptionsUnsqueezeShapeuse_deterministic_algorithmsValidateBatchedSquareMatrixValidateChunkDimValidatedEinsumShapevalidateDeviceValidateDeviceValidatedRearrangeShapeValidatedReduceShapeValidatedRepeatShapevalidateDTypeValidateEinsumValidateOperandCountValidateRanksValidateScalarValidateSplitDimValidateSquareMatrixValidateUnbindDimValueOptionsvar_var_meanvdotviewview_as_complexview_as_realvmapvsplitvstackWebGPUKernelConfigWebGPUOnlyResultWebGPUTensorDatawhereWindowOptionsxlogyzeroszeros_like
torch.js· 2026
LegalTerms of UsePrivacy Policy
/
/
  1. docs
  2. torch.js
  3. torch
  4. nn
  5. TransformerDecoder

torch.nn.TransformerDecoder

class TransformerDecoder extends Module
new TransformerDecoder(options: TransformerDecoderOptions)
readonlynum_layers(number)
layers(ModuleList)
norm(LayerNorm | null)

Stack of N Transformer decoder layers.

A TransformerDecoder applies multiple TransformerDecoderLayer sequentially, with each layer processing the target sequence while attending to the encoder memory. This enables auto-regressive sequence generation and encoder-decoder sequence-to-sequence models.

Key differences from TransformerEncoder:

  • Each layer has cross-attention to encoder output (memory), not just self-attention
  • Usually used with causal masking on self-attention to prevent attending to future tokens
  • Designed for generation tasks where tokens are produced one at a time

Commonly used in:

  • Machine translation (encoder-decoder models like BERT2BERT, mBART)
  • Image captioning (visual encoder + text decoder)
  • Sequence-to-sequence models (speech-to-text, abstractive summarization)
  • Auto-regressive text generation (with causal masking)
  • Question answering (encode question, decode answer)
textoutput=textLayerN(...textLayer2(textLayer1(tgt),memory)...,memory)textwithoptionalfinalnorm:output=textLayerNorm(textoutput)\begin{aligned} \\text{output} = \\text{Layer}_N(...\\text{Layer}_2(\\text{Layer}_1(tgt), memory)..., memory) \\ \\text{with optional final norm: output} = \\text{LayerNorm}(\\text{output}) \end{aligned}textoutput=textLayerN​(...textLayer2​(textLayer1​(tgt),memory)...,memory)textwithoptionalfinalnorm:output=textLayerNorm(textoutput)​
  • Cross-attention to memory: Each decoder layer attends to the encoder output (memory). Memory shape is [src_seq_len, batch, d_model], typically different from target sequence length.
  • Causal masking: When generating auto-regressively, use causal mask on self-attention to prevent attending to future tokens. During training with teacher forcing, you can optionally use it too.
  • Final layer normalization: Including norm (typically LayerNorm(d_model)) stabilizes outputs. Highly recommended, especially for deep decoders.
  • Target and memory shapes: Target has shape [tgt_seq_len, batch, d_model], memory has shape [src_seq_len, batch, d_model]. Sequences can be different lengths!
  • Depth vs encoder: Decoder often has same depth as encoder (both 6 layers in base Transformer), but can differ. Deeper decoders (12+ layers) used in large language models.
  • Computational complexity: O(num_layers × tgt_seq_len × (tgt_seq_len + src_seq_len) × d_model). Depends on both target and source lengths.
  • Causal mask MUST be applied to self-attention (not cross-attention). Cross-attention can freely attend to memory.
  • Memory must come from the same encoder (usually). If encoder changes, results will differ significantly.
  • During inference with auto-regression, typically use growing target sequence (KV caching is not implemented).
  • Ensure tgt has shape [tgt_seq_len, batch, d_model] and memory has shape [src_seq_len, batch, d_model].
  • Padding masks should indicate which positions are padding (True = ignore), not which are valid.

Examples

// Create decoder with 6 layers
const decoder_layer = new torch.nn.TransformerDecoderLayer({
  d_model: 512,
  nhead: 8,
  dim_feedforward: 2048,
  dropout: 0.1
});

const decoder = new torch.nn.TransformerDecoder({
  decoder_layer,
  num_layers: 6,
  norm: new torch.nn.LayerNorm(512)
});

// Get memory from encoder
const src = torch.randn(10, 32, 512);
const memory = encoder.encode(src);  // [seq_len, batch, d_model]

// Decode from target to output
const tgt = torch.randn(20, 32, 512);  // target sequence [seq_len, batch, d_model]
const output = decoder.decode(tgt, memory);

// With causal mask for auto-regressive generation
const tgt_mask = torch.nn.Transformer.generate_square_subsequent_mask(20);
const output = decoder.decode(tgt, memory, { tgt_mask });

// Full encoder-decoder pipeline
const encoded = encoder.encode(src);
const decoded = decoder.decode(tgt, encoded);
const predictions = output_projection(decoded);  // project to vocab

// During inference: generate one token at a time
let generated = [bos_token_id];
for (let step = 0; step < max_length; step++) {
  const tgt_emb = embedding(torch.tensor(generated));
  const tgt_mask = torch.nn.Transformer.generate_square_subsequent_mask(generated.length);
  const out = decoder.decode(tgt_emb, memory, { tgt_mask });
  const logits = output_proj(out[-1]);  // last token
  const next_token = torch.argmax(logits).item();
  generated.push(next_token);
  if (next_token === eos_token_id) break;
}

See Also

  • PyTorch torch.nn.TransformerDecoder
Previous
Transformer.run
Next
TransformerDecoder.decode