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. MultiheadAttention
  6. multihead_attn

torch.nn.MultiheadAttention.multihead_attn

MultiheadAttention.multihead_attn(query: Tensor, options?: MultiheadAttnOptions): [Tensor, Tensor | null]MultiheadAttention.multihead_attn(query: Tensor, key: Tensor, value: Tensor, key_padding_mask: Tensor | undefined, need_weights: boolean, attn_mask: Tensor | undefined, average_attn_weights: boolean, is_causal: boolean, options?: MultiheadAttnOptions): [Tensor, Tensor | null]

Full multi-head attention with comprehensive control over masking and outputs.

This method provides complete control over attention computation including support for causal masking (autoregressive generation), padding masks (variable-length sequences), custom attention masks, and returning attention weights for interpretability. It's the main entry point for all attention patterns: self-attention, cross-attention (encoder-decoder), and masked attention for generation.

Computation flow:

  1. Project query, key, value tensors using learned weight matrices
  2. Reshape embeddings into separate attention heads (num_heads parallel subspaces)
  3. Apply scaled dot-product attention for each head: softmax((QK^T)/√d_k) V
  4. Apply masks (causal, padding, or custom) before softmax to control which positions can attend
  5. Apply dropout to attention weights for regularization (training mode only)
  6. Concatenate outputs from all heads and apply output projection
  7. Optionally return attention weights (averaged across heads or per-head)

Input shapes: (seq_len, batch, embed_dim) or (batch, seq_len, embed_dim) if batch_first

  • L: target sequence length (query length)
  • S: source sequence length (key/value length, can differ from L for cross-attention)
  • N: batch size
  • E: embedding dimension (embed_dim)
  • d_k: head dimension = embed_dim / num_heads

Output shapes: Same as query input, or transposed to batch-first if needed

Attention(Q,K,V)=softmax(QKTdk+M)Vdk=embed_dimnum_headsM=attention mask (additive, -∞ for masked positions)\begin{aligned} \text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}} + M\right) V \\ d_k = \frac{\text{embed\_dim}}{\text{num\_heads}} \\ M = \text{attention mask (additive, -∞ for masked positions)} \end{aligned}Attention(Q,K,V)=softmax(dk​​QKT​+M)Vdk​=num_headsembed_dim​M=attention mask (additive, -∞ for masked positions)​
  • Shape requirements: query and key must have compatible batch dimensions. Query length (L) determines output length. Key/value length (S) determines which positions can attend.
  • Causal vs attn_mask: is_causal=true and attn_mask cannot be used together. Use is_causal for autoregressive models, attn_mask for custom patterns.
  • Mask syntax: Boolean masks (key_padding_mask) where True=masked. Additive masks (attn_mask) where negative values suppress attention. Both get combined additively.
  • Dropout during eval: Dropout is disabled during evaluation mode (self.training=false). Always call model.eval() before inference to match training behavior.
  • Head dimension: embed_dim must be divisible by num_heads. Each head processes embed_dim/num_heads dimensions. This is validated in constructor, but useful to know for debugging.
  • Attention weights interpretation: High values indicate strong attention. weights[b, q, k] shows how much position q attends to position k. Averaged across heads gives overall pattern.
  • Numerical stability: Scaled dot-product attention divides by √d_k to prevent softmax saturation. Masking is applied before softmax for numerical stability (-inf instead of post-softmax).

Parameters

queryTensor
Query tensor of shape (L, N, E) or (N, L, E) if batch_first. Determines what to attend to. Length L determines output length. For self-attention, use same tensor as key/value.
optionsMultiheadAttnOptionsoptional
Optional dictionary controlling attention behavior: Masking options: - attn_mask - Custom attention mask of shape (L, S) or (N*num_heads, L, S). True/1.0 positions are masked (set to -inf before softmax). Used for custom patterns beyond causal/padding. - key_padding_mask - Boolean mask of shape (N, S) where True marks padded positions in key. Prevents attention to padding tokens. Essential for batches with variable lengths. - is_causal - If true, applies causal mask restricting each position from attending to future positions. Essential for autoregressive generation (language models, next-token prediction). Cannot be combined with attn_mask. Enforces: position i only attends to positions [0...i]. Output options: - need_weights - If true, returns attention weights as second tuple element. Default: true. Weights are useful for visualization and interpretability but add computation cost. - average_attn_weights - If true (default), averages weights across all heads to produce shape (N, L, S) for single attention weight matrix. If false, returns per-head weights with shape (N, num_heads, L, S) - useful for analysis.

Returns

[Tensor, Tensor | null]– Tuple of (attention_output, attention_weights): - attention_output: Tensor of shape (L, N, E) or (N, L, E) if batch_first. Contextual representation where each position has attended to relevant positions weighted by attention scores. - attention_weights: Tensor of shape (N, L, S) if average_attn_weights=true, or (N, num_heads, L, S) if average_attn_weights=false. Null if need_weights=false. Values in [0, 1] indicating how much each position attended to each key position.

Examples

// Causal self-attention for next-token prediction
const attn = new torch.nn.MultiheadAttention({
  embed_dim: 512,
  num_heads: 8,
  dropout: 0.1,
  batch_first: true
});

const x = torch.randn([batch_size, seq_len, 512]);
const [output, weights] = attn.multihead_attn(x, x, x, {
  is_causal: true,  // Only attend to past positions
  need_weights: true
});

// weights[i, j, k] = attention from position j to position k
// For causal: weights[i, j, k] = 0 for k > j (future positions masked)
// Cross-attention with encoder-decoder
const attn = new torch.nn.MultiheadAttention({
  embed_dim: 768,
  num_heads: 12,
  kdim: 768,
  vdim: 768,
  batch_first: true
});

const encoder_output = torch.randn([batch_size, src_len, 768]);  // Encoder outputs
const decoder_hidden = torch.randn([batch_size, tgt_len, 768]);  // Decoder hidden state

// Decoder attends to encoder outputs
const [context, weights] = attn.multihead_attn(
  decoder_hidden,     // query: what to attend to
  encoder_output,     // key/value: encoder outputs
  encoder_output,
  { need_weights: false }
);
// context: decoder representations augmented with encoder information
// Variable-length batch with padding mask
const attn = new torch.nn.MultiheadAttention({
  embed_dim: 256,
  num_heads: 4,
  batch_first: true
});

const max_len = 100;
const batch = torch.randn([32, max_len, 256]);
const actual_lens = torch.tensor([80, 95, 100, 75, ...]);  // 32 values

// Create padding mask
const positions = torch.arange(max_len).unsqueeze(0);
const padding_mask = positions >= actual_lens.unsqueeze(1);  // [32, max_len]

const [output, weights] = attn.multihead_attn(
  batch, batch, batch,
  { key_padding_mask: padding_mask }
);
// Padded positions contribute 0 attention weight to all positions

See Also

  • forward - Simplified self-attention interface without options
  • torch.nn.functional.scaled_dot_product_attention - Core attention computation
  • torch.nn.TransformerEncoderLayer - Complete layer with attention + feedforward
  • torch.nn.TransformerDecoderLayer - Decoder layer with self+cross attention
Previous
MultiheadAttention.load_state_dict
Next
MultiheadAttention.state_dict