Пример #1
0
def ConfigurableTerraformer(input_vocab_size,
                            output_vocab_size=None,
                            d_model=512,
                            d_ff=2048,
                            d_attention_key=None,
                            d_attention_value=None,
                            n_encoder_layers=6,
                            n_decoder_layers=6,
                            n_heads=8,
                            dropout=0.1,
                            max_len=2048,
                            encoder_attention_type=tl.SelfAttention,
                            encoder_decoder_attention_type=tl.SelfAttention,
                            pos_type='fixed-base',
                            pos_axial_shape=(),
                            pos_d_axial_embs=None,
                            pos_start_from_zero_prob=1.0,
                            pos_max_offset_to_add=0,
                            ff_activation=tl.Relu,
                            ff_use_sru=0,
                            ff_chunk_size=0,
                            ff_dropout=None,
                            ff_sparsity=0,
                            loss_sparsity_type='mult',
                            loss_sparsity=0,
                            loss_d_lowrank=0,
                            loss_sparsity_prob=None,
                            attention_chunk_size=0,
                            n_layers_forget=0,
                            forget_dense=True,
                            n_decoder_attention_layers=2,
                            use_bfloat16=False,
                            reversible_encoder=False,
                            use_two_swaps_per_encoder_block=True,
                            center_layernorm=True,
                            half_before_layer=None,
                            double_after_layer=None,
                            mode='train'):
    """Returns a highly configurable Terraformer encoder-decoder model.

  This model maps paired text sequences (source and target) to float-valued
  losses. If ``input_vocab_size`` is not ``None``, the layer takes
  two input sequences:

    - inputs (2):

        - source: 2-D int array representing a batch of text strings via token
          IDs plus padding markers; shape is `(batch_size, sequence_length)`,
          where sequence_length <= ``max_len``. Array elements are in
          ``range(input_vocab_size)``, and 0 values mark padding positions.

        - target: 2-D int array representing a batch of text strings via token
          IDs plus padding markers; shape is `(batch_size, sequence_length)`,
          where sequence_length <= ``max_len``. Array elements are in
          ``range(output_vocab_size)``, and 0 values mark padding positions.

    - output: 1-D float array of losses; shape is `(batch_size)`.

  If ``input_vocab_size`` is ``None``, the layer takes three input sequences:

    - inputs (3):

        - source: 3-D float array representing a batch of already-embedded text
          strings; shape is `(batch_size, sequence_length, d_model)`, where
          sequence_length <= ``max_len``.

        - mask: 2-D int array representing active versus masked positions; 0
          values mark masked (padding) positions.

        - target: 2-D int array representing a batch of text strings via token
          IDs plus padding markers; shape is `(batch_size, sequence_length)`,
          where sequence_length <= ``max_len``. Array elements are in
          ``range(output_vocab_size)``, and 0 values mark padding positions.

    - output: 1-D float array of losses; shape is `(batch_size)`.

  Args:
    input_vocab_size: Input vocabulary size -- each element of the input tensor
        should be an integer in ``range(vocab_size)``. These integers typically
        represent token IDs from a vocabulary-based tokenizer.
    output_vocab_size: If specified, gives the vocabulary size for the targets;
        if ``None``, then input and target integers (token IDs) are assumed to
        come from the same vocabulary.
    d_model: Last/innermost dimension of activation arrays at most points in
        the model, including the initial embedding output.
    d_ff: Last/innermost dimension of special (typically wider)
        :py:class:`Dense` layer in the feedforward part of each encoder block.
    d_attention_key: Depth of key vectors in each attention head.
    d_attention_value: Depth of value vectors in each attention head.
    n_encoder_layers: Number of encoder blocks.
    n_decoder_layers: Number of decoder blocks.
    n_heads: Number of attention heads.
    dropout: Stochastic rate (probability) for dropping an activation value
        when applying dropout within encoder/decoder blocks. The same rate is
        also used for attention dropout in encoder/decoder blocks.
    max_len: Maximum symbol length for positional encoding.
    encoder_attention_type: Type of attention to use in the encoder; must be
        an attention-type subclass of :py:class:`trax.layers.Layer`.
    encoder_decoder_attention_type: Type of attention to use in the decoder;
        must be an attention-type subclass of :py:class:`trax.layers.Layer`.
    pos_type: String indicating the type of positional embeddings to use.
    pos_axial_shape: Shape (tuple of ints) to use for the axial position
      encoding. If unset, axial position encoding is disabled.
    pos_d_axial_embs: Tuple of ints specifying the depth of position embedding
        for each axis. Tuple length must match ``pos_axial_shape``, and values
        must sum to ``d_model``.
    pos_start_from_zero_prob: Stochastic rate (probability) for starting
        positional encoding at position 0 during training. If 1.0, always start
        from position 0; if < 1.0, the non-zero starts will be uniformly
        distributed up to ``pos_max_offset_to_add``.
    pos_max_offset_to_add: Maximum offset to add to positions during training
        when randomizing. This offset plus input length must be less than
        ``max_len`` for all training examples.
    ff_activation: Type of activation function at the end of each block; must
        be an activation-type subclass of :py:class:`trax.layers.Layer`.
    ff_use_sru: If > 0, use this number of SRU layers in place of feedforward
        layers.
    ff_chunk_size: If > 0, chunk each feedforward layer into chunks of this
        size.
    ff_dropout: Stochastic rate (probability) for dropping an activation value
        at feedforward nonlinearities.
    ff_sparsity: If > 0, use sparse feedforward blocks with this level of
        sparsity.
    loss_sparsity_type: String indicating the type of sparsity to used in loss
        layer; see :py:class:`SparseDenseWithOptions` for options. If ``None``,
        use no sparsity.
    loss_sparsity: If > 0, use this level of sparsity in the loss layer.
    loss_d_lowrank: If > 0, use a (low-rank) intermediate layer, with this
        dimension, in the loss.
    loss_sparsity_prob: Stochastic rate (probability) for using the sparse
        version of the loss. If ``None``, use the sparse version exclusively.
    attention_chunk_size: If > 0, compute attention using chunks of this size.
    n_layers_forget: How often to have a forgetting block between layers.
    forget_dense: If True, use :py:class:`Dense` instances as forget layers;
        else use no-ops.
    n_decoder_attention_layers: Number of attention layers in a decoder block.
    use_bfloat16: If True, use bfloat16 for weights; else use float32.
    reversible_encoder: If True, make the encoder be reversible.
    use_two_swaps_per_encoder_block: If True, ensure that there is a an even
        number of swaps across the encoder.
    center_layernorm: If True, use centering in :py:class:`LayerNorm` (the
        default); else omit centering (which is known as RMS normalization).
    half_before_layer: If not None, specifies an n'th layer such that all
        layers before the n'th use half the normal values for ``d_model`` and
        ``d_ff``.
    double_after_layer: If not None, specifies an n'th layer such that all
        layers after the n'th use double the normal values for ``d_model`` and
        ``d_ff``.
    mode: If ``'train'``, include dropout in each encoder/decoder block; else
        dropout layers have no effect.

  Returns:
    A Terraformer encoder-decoder as a layer that maps from target and source
    text sequences to a scalar loss.
  """
    if mode == 'predict':
        portal_mask = _PortalInput()
    else:
        portal_mask = None

    # Set default dimensions for attention head key and value sizes.
    if (d_model / 2) % n_heads != 0:
        raise ValueError(
            f'n_heads ({n_heads}) must divide d_model/2 ({d_model/2})')
    if d_attention_key is None:
        d_attention_key = d_model // n_heads
    if d_attention_value is None:
        d_attention_value = d_model // n_heads

    # Set values of d_model, d_ff and d_qkv for the first stage.
    d_model1, d_ff1 = d_model, d_ff
    d_attention_key1, d_attention_value1 = d_attention_key, d_attention_value
    if half_before_layer:
        d_model1, d_ff1 = d_model / 2, d_ff / 2
        d_attention_key1 = d_attention_key / 2
        d_attention_value1 = d_attention_value / 2

    # Set values of d_model, d_ff and d_qkv for the final stage.
    d_model2, d_ff2 = d_model, d_ff
    d_attention_key2, d_attention_value2 = d_attention_key, d_attention_value
    if double_after_layer:
        d_model2, d_ff2 = d_model * 2, d_ff * 2
        d_attention_key2 = d_attention_key * 2
        d_attention_value2 = d_attention_value * 2

    # Vector embeddings.
    in_encoder, out_encoder, output_vocab_size = (
        ct.EmbeddingAndPositionalEncodings(
            input_vocab_size,
            d_model1,
            mode,
            dropout,
            [-2],  # dropout_shared_axes
            max_len,
            output_vocab_size=output_vocab_size,
            pos_type=pos_type,
            pos_axial_shape=pos_axial_shape,
            pos_d_axial_embs=pos_d_axial_embs,
            pos_start_from_zero_prob=pos_start_from_zero_prob,
            pos_max_offset_to_add=pos_max_offset_to_add,
            use_bfloat16=use_bfloat16))

    def _EncoderBlock():
        return reformer.EncoderBlock(
            d_model1,
            d_ff1,
            n_heads,
            encoder_attention_type,
            dropout=dropout,
            ff_activation=ff_activation,
            ff_dropout=ff_dropout,
            ff_use_sru=ff_use_sru,
            ff_chunk_size=ff_chunk_size,
            ff_sparsity=ff_sparsity,
            attention_chunk_size=attention_chunk_size,
            center_layernorm=center_layernorm,
            use_bfloat16=use_bfloat16,
            use_two_swaps_per_block=use_two_swaps_per_encoder_block,
            mode=mode)

    def _Encoder():  # vec_e mask_e tok_e tok_d tok_d
        layers = [
            tl.ReversibleSelect([0, 0]),
            _ReversibleSerialForget(
                [_EncoderBlock() for _ in range(n_encoder_layers)], d_model1,
                n_layers_forget, forget_dense)
        ]
        if not reversible_encoder:
            layers += [
                _XYAvg(),
                tl.Dense(d_model1, use_bfloat16=use_bfloat16),
                tl.LayerNorm(),
            ]
        if mode == 'predict':
            return tl.Cache(tl.Serial(layers))
        else:
            return tl.Serial(layers)

    if mode == 'predict':
        # TODO(jaszczur): Remove temporary fix of Terraformer padding in predict.
        # In predict mode Terraformer needs masking for merged encoder-decoder
        # sequence. This monkey patches the layer with a mask to neccessary places.
        # This shouldn't be a permanent solution - mask should be passed through
        # the stack and all the layers.
        tl.attention.DotProductCausalAttention.monkey_patched_mask = (
            lambda x: portal_mask)
        tl.research.sparsity._RememberPad.monkey_patched_mask = (  # pylint: disable=protected-access
            lambda x: portal_mask)
        originalScanSRUCell = tl.rnn.ScanSRUCell
        tl.rnn.ScanSRUCell = functools.partial(tl.rnn.ScanSRUCell,
                                               monkey_patched_mask=portal_mask)

    decoder_blocks = []

    if isinstance(encoder_decoder_attention_type, (tuple, list)):
        assert n_decoder_layers % len(encoder_decoder_attention_type) == 0
    else:
        encoder_decoder_attention_type = [encoder_decoder_attention_type]
    for layer_idx in range(n_decoder_layers):
        layer_attention_type = encoder_decoder_attention_type[
            layer_idx % len(encoder_decoder_attention_type)]
        # Grow d_model, d_ff, and d_qkv if requested.
        d_m, d_f, d_k, d_v = d_model1, d_ff1, d_attention_key1, d_attention_value1
        if half_before_layer and layer_idx >= half_before_layer:
            d_m, d_f, d_k, d_v = d_model, d_ff, d_attention_key, d_attention_value
        if double_after_layer and layer_idx > double_after_layer:
            d_m, d_f, d_k, d_v = d_model2, d_ff2, d_attention_key2, d_attention_value2
        decoder_block = reformer.DecoderBlock(
            d_m,
            d_f,
            d_k,
            d_v,
            n_heads,
            attention_type=layer_attention_type,
            dropout=dropout,
            ff_activation=ff_activation,
            ff_dropout=ff_dropout,
            ff_use_sru=ff_use_sru,
            ff_chunk_size=ff_chunk_size,
            ff_sparsity=ff_sparsity,
            attention_chunk_size=attention_chunk_size,
            n_attention_layers=n_decoder_attention_layers,
            center_layernorm=center_layernorm,
            use_bfloat16=use_bfloat16,
            mode=mode)
        decoder_blocks.append(decoder_block)
        if half_before_layer and layer_idx == half_before_layer - 1:
            decoder_blocks.append(tl.ReversibleConcatenatePair())
        if double_after_layer and layer_idx == double_after_layer:
            decoder_blocks.append(tl.ReversibleConcatenatePair())

    if mode == 'predict':
        # After initializing the decoder we can revert to original state of
        # previously monkey-patched classes/functions.
        tl.attention.DotProductCausalAttention.monkey_patched_mask = (
            lambda x: None)
        tl.research.sparsity._RememberPad.monkey_patched_mask = (lambda x: None
                                                                 )  # pylint: disable=protected-access
        tl.rnn.ScanSRUCell = originalScanSRUCell

    def _Loss():
        return tl.SparseDenseWithOptions(output_vocab_size,
                                         d_input=d_model2,
                                         sparsity_type=loss_sparsity_type,
                                         sparsity=loss_sparsity,
                                         d_lowrank=loss_d_lowrank,
                                         prob_sparse=loss_sparsity_prob,
                                         use_bfloat16=use_bfloat16,
                                         mode=mode)

    def _enc_dec_concat():
        """Layers to merge encoder and decoder."""
        if reversible_encoder:
            return [
                tl.ReversibleSelect([0, 1, 4, 2,
                                     3]),  # v_e v_d mask_e tok_e tok_d
                t2.ConcatWithPadding2(mode=mode),  # v_ed v_ed tok_e tok_d
            ]
        else:
            return [
                tl.ReversibleSelect([0, 3, 1,
                                     2]),  # v_e v_d mask_e tok_e tok_d
                t2.ConcatWithPadding(mode=mode),  # v_ed tok_e tok_d
                tl.ReversibleSelect([0, 0]),  # v_ed v_ed tok_e tok_d
            ]

    def _inp_layers():
        if input_vocab_size is not None:
            return tl.AssertFunction(
                'bl,br->bld,bl,bl,br',  # b: batch, l/r: enc/dec length, d: vec depth
                tl.Serial(  # tok_e tok_d
                    tl.Select([0, 0, 0, 1]),
                    tl.Parallel(
                        in_encoder,
                        [tl.PaddingMask(), _RemoveAxes12()
                         ])))  # vec_e mask_e tok_e tok_d
        else:
            # Input in this case is vec_e, mask_e, tok_d. Where all downstream
            # operations expect tok_e, we give it instead mask_e, expecting that
            # downstream ops only are looking for padding/not padding.
            return tl.AssertFunction(
                'blf,bl,br->bld,bl,bl,br',  # f: in-feature depth, d: out-vector depth
                tl.Serial(  # vec_e mask_e tok_d
                    tl.Select([0, 1, 1, 2]),
                    tl.Parallel(in_encoder, [],
                                _AsTokenIDs())))  # vec_e mask_e tok_e tok_d

    # Assemble and return the model.
    return tl.Serial(
        _inp_layers(),  # vec_e mask_e tok_e tok_d
        tl.Parallel([], portal_mask),
        tl.Select([0, 1, 2, 3, 3]),  # Copy decoder tokens for use in loss.

        # Embed in and out tokens; done together as weights may be shared.
        tl.Parallel([], [], [], [tl.ShiftRight(mode=mode), out_encoder
                                 ]),  # vec_e mask_e tok_e vec_d tok_d

        # Encode; then concat encoder and decoder, given encoder mask.
        _Encoder(),  # vec_e mask_e tok_e vec_d tok_d
        _enc_dec_concat(),

        # Run decoder blocks.
        _ReversibleSerialForget(decoder_blocks, d_model2, n_layers_forget,
                                forget_dense),  # vec_ed1 vec_ed2 tok_e tok_d
        _XYAvg(),  # vec_ed tok_e tok_d
        tl.LayerNorm(),  # vec_ed tok_e tok_d

        # Separate out the encoder part from the concatenated vector,
        # then compute loss.
        tl.Select([0, 1, 2, 2]),  # vec_ed tok_e tok_d tok_d
        t2.StripFromConcatenateWithPadding(mode=mode),  # vec_d tok_d
        _Loss(),  # vec_d tok_d
    )
Пример #2
0
def Reformer2(input_vocab_size,
              output_vocab_size=None,
              d_model=512,
              d_ff=2048,
              d_attention_key=None,
              d_attention_value=None,
              n_encoder_layers=6,
              n_decoder_layers=6,
              n_heads=8,
              dropout=0.1,
              max_len=2048,
              encoder_attention_type=tl.SelfAttention,
              encoder_decoder_attention_type=tl.SelfAttention,
              pos_type='fixed-base',
              pos_axial_shape=(),
              pos_d_axial_embs=None,
              pos_start_from_zero_prob=1.0,
              pos_max_offset_to_add=0,
              ff_activation=tl.Relu,
              ff_use_sru=0,
              ff_chunk_size=0,
              ff_dropout=None,
              ff_sparsity=0,
              loss_sparsity_type='mult',
              loss_sparsity=0,
              loss_d_lowrank=0,
              loss_sparsity_prob=None,
              attention_chunk_size=0,
              n_layers_forget=0,
              forget_dense=True,
              n_decoder_attention_layers=2,
              use_bfloat16=False,
              reversible_encoder=False,
              use_two_swaps_per_encoder_block=True,
              center_layernorm=True,
              half_before_layer=None,
              double_after_layer=None,
              mode='train'):
    """Reversible transformer encoder-decoder model.

  This model expects an input pair: source, target.

  At the moment, this model supports dot-product attention only. For the
  attention types in the Reformer paper, see ReformerLM.

  Args:
    input_vocab_size: int: vocab size of the source.
    output_vocab_size: int (optional): vocab size of the target. If None, the
      source and target are assumed to have the same vocab.
    d_model: int:  depth of embedding
    d_ff: int: depth of feed-forward layer
    d_attention_key: int: depth of key vector for each attention head
    d_attention_value: int: depth of value vector for each attention head
    n_encoder_layers: int: number of encoder layers
    n_decoder_layers: int: number of decoder layers
    n_heads: int: number of attention heads
    dropout: float: dropout rate (how much to drop out)
    max_len: int: maximum symbol length for positional encoding
    encoder_attention_type: class: attention class to use, such as SelfAttention
    encoder_decoder_attention_type: class: attention class to use, such as
      SelfAttention
    pos_type: string, the type of positional embeddings to use.
    pos_axial_shape: tuple of ints: input shape to use for the axial position
      encoding. If unset, axial position encoding is disabled.
    pos_d_axial_embs: tuple of ints: depth of position embedding for each axis.
      Tuple length must match pos_axial_shape, and values must sum to d_model.
    pos_start_from_zero_prob: how often to start from 0 during training,
          (if 1.0, we always start from position 0, if less, we randomize).
    pos_max_offset_to_add: maximum offset to add to positions during training
        when randomizing; this offset plus input length must still be less than
        max_len for all training examples.
    ff_activation: the non-linearity in feed-forward layer
    ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
    ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
    ff_dropout: float: (optional) separate dropout rate at feed-forward
      nonlinearity. This is called relu_dropout in T2T.
    ff_sparsity: int, if > 0 use sparse feed-forward block with this sparsity
    loss_sparsity_type: str, type of sparsity to used in loss layer. See
      SparseDenseWithOptions for options. None if no sparsity should be used.
    loss_sparsity: int, the sparsity for loss layer (if used)
    loss_d_lowrank: int, the dimensions for intermediate layer (if used)
    loss_sparsity_prob: float, the probability for sparse version of loss to be
      used. If None, only sparse version is used.
    attention_chunk_size: int, if > 0 run attention chunked at this size
    n_layers_forget: how often to have a forgetting block between layers
    forget_dense: whether to use Dense or no-op (Serial) as a forget layer.
    n_decoder_attention_layers: how many attention layers in a decoder block
    use_bfloat16: whether to use bfloat16 for weights (default: False)
    reversible_encoder: whether to be reversible through the encoder
    use_two_swaps_per_encoder_block: whether to allow even number of swaps in
      the encoder
    center_layernorm: whether to use centering in LayerNorm (default) or if
      to skip it, which is known as RMS normalization.
    half_before_layer: int, half d_model and d_ff before that layer
    double_after_layer: int, double d_model and d_ff after that layer
    mode: str: 'train' or 'eval'

  Returns:
    A Reformer model as a layer that maps from a target, source pair to
    activations over a vocab set.
  """
    # Set default dimensions for attention head key and value sizes.
    if (d_model / 2) % n_heads != 0:
        raise ValueError(
            f'n_heads ({n_heads}) must divide d_model/2 ({d_model/2})')
    if d_attention_key is None:
        d_attention_key = d_model // n_heads
    if d_attention_value is None:
        d_attention_value = d_model // n_heads

    # Set values of d_model, d_ff and d_qkv for the first stage.
    d_model1, d_ff1 = d_model, d_ff
    d_attention_key1, d_attention_value1 = d_attention_key, d_attention_value
    if half_before_layer:
        d_model1, d_ff1 = d_model / 2, d_ff / 2
        d_attention_key1 = d_attention_key / 2
        d_attention_value1 = d_attention_value / 2

    # Set values of d_model, d_ff and d_qkv for the final stage.
    d_model2, d_ff2 = d_model, d_ff
    d_attention_key2, d_attention_value2 = d_attention_key, d_attention_value
    if double_after_layer:
        d_model2, d_ff2 = d_model * 2, d_ff * 2
        d_attention_key2 = d_attention_key * 2
        d_attention_value2 = d_attention_value * 2

    # Vector embeddings.
    in_encoder, out_encoder, output_vocab_size = (
        ct.EmbeddingAndPositionalEncodings(
            input_vocab_size,
            d_model1,
            mode,
            dropout,
            [-2],  # dropout_shared_axes
            max_len,
            output_vocab_size=output_vocab_size,
            pos_type=pos_type,
            pos_axial_shape=pos_axial_shape,
            pos_d_axial_embs=pos_d_axial_embs,
            pos_start_from_zero_prob=pos_start_from_zero_prob,
            pos_max_offset_to_add=pos_max_offset_to_add,
            use_bfloat16=use_bfloat16))

    # pylint: disable=g-complex-comprehension
    encoder_blocks = [
        EncoderBlock(d_model1,
                     d_ff1,
                     n_heads,
                     encoder_attention_type,
                     dropout=dropout,
                     ff_activation=ff_activation,
                     ff_dropout=ff_dropout,
                     ff_use_sru=ff_use_sru,
                     ff_chunk_size=ff_chunk_size,
                     ff_sparsity=ff_sparsity,
                     attention_chunk_size=attention_chunk_size,
                     center_layernorm=center_layernorm,
                     use_bfloat16=use_bfloat16,
                     use_two_swaps_per_block=use_two_swaps_per_encoder_block,
                     mode=mode) for _ in range(n_encoder_layers)
    ]
    # pylint: enable=g-complex-comprehension

    encoder = [  # vec_e mask_e tok_e tok_d tok_d
        tl.ReversibleSelect([0, 0]),  # vec_e1 vec_e2 mask_e tok_e tok_d tok_d
        _ReversibleSerialForget(encoder_blocks, d_model1, n_layers_forget,
                                forget_dense)
    ]
    if not reversible_encoder:
        encoder += [
            tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0),
            tl.Dense(d_model1, use_bfloat16=use_bfloat16),
            tl.LayerNorm(),
        ]
    encoder = tl.Serial(encoder)
    if mode == 'predict':
        encoder = tl.Cache(encoder)

    decoder_blocks = []

    if isinstance(encoder_decoder_attention_type, (tuple, list)):
        assert n_decoder_layers % len(encoder_decoder_attention_type) == 0
    else:
        encoder_decoder_attention_type = [encoder_decoder_attention_type]
    for layer_idx in range(n_decoder_layers):
        layer_attention_type = encoder_decoder_attention_type[
            layer_idx % len(encoder_decoder_attention_type)]
        # Grow d_model, d_ff, and d_qkv if requested.
        d_m, d_f, d_k, d_v = d_model1, d_ff1, d_attention_key1, d_attention_value1
        if half_before_layer and layer_idx >= half_before_layer:
            d_m, d_f, d_k, d_v = d_model, d_ff, d_attention_key, d_attention_value
        if double_after_layer and layer_idx > double_after_layer:
            d_m, d_f, d_k, d_v = d_model2, d_ff2, d_attention_key2, d_attention_value2
        decoder_block = DecoderBlock(
            d_m,
            d_f,
            d_k,
            d_v,
            n_heads,
            attention_type=layer_attention_type,
            dropout=dropout,
            ff_activation=ff_activation,
            ff_dropout=ff_dropout,
            ff_use_sru=ff_use_sru,
            ff_chunk_size=ff_chunk_size,
            ff_sparsity=ff_sparsity,
            attention_chunk_size=attention_chunk_size,
            n_attention_layers=n_decoder_attention_layers,
            center_layernorm=center_layernorm,
            use_bfloat16=use_bfloat16,
            mode=mode)
        decoder_blocks.append(decoder_block)
        if half_before_layer and layer_idx == half_before_layer - 1:
            decoder_blocks.append(tl.ReversibleConcatenatePair())
        if double_after_layer and layer_idx == double_after_layer:
            decoder_blocks.append(tl.ReversibleConcatenatePair())

    dense_loss_layer = tl.SparseDenseWithOptions(
        output_vocab_size,
        d_input=d_model2,
        sparsity_type=loss_sparsity_type,
        sparsity=loss_sparsity,
        d_lowrank=loss_d_lowrank,
        prob_sparse=loss_sparsity_prob,
        use_bfloat16=use_bfloat16,
        mode=mode)

    # Layers to merge encoder and decoder, see below for details.
    if reversible_encoder:
        encdec_layers = [
            tl.ReversibleSelect([0, 1, 4, 2,
                                 3]),  # vec_e vec_d mask_e tok_e tok_d
            t2.ConcatWithPadding2(mode=mode),  # vec_ed vec_ed tok_e tok_d
        ]
    else:
        encdec_layers = [
            tl.ReversibleSelect([0, 3, 1,
                                 2]),  # vec_e vec_d mask_e tok_e tok_d
            t2.ConcatWithPadding(mode=mode),  # vec_ed tok_e tok_d
            tl.ReversibleSelect([0, 0]),  # vec_ed vec_ed tok_e tok_d
        ]

    # Assemble and return the model.
    return tl.Serial(
        # Input: encoder_side_tokens, decoder_side_tokens
        # Copy decoder tokens for use in loss.
        tl.Select([0, 0, 0, 1, 1]),  # tok_e tok_e tok_e tok_d tok_d

        # Embed in and out tokens; done together as weights may be shared.
        tl.Parallel(
            in_encoder,
            [],
            [],  # vec_e tok_e tok_e vec_d tok_d
            [tl.ShiftRight(mode=mode), out_encoder]),

        # Predict mode doesn't work with padding in encoder. Raising an exception
        # in jitted function isn't possible, so the second next best thing is
        # to convert every embedding to NaNs, so the user will not get subtly
        # wrong results, but clearly wrong results.
        (_ConvertToNaNsOnAnyZero() if mode == 'predict' else []),
        tl.Parallel([], [
            tl.PaddingMask(),
            tl.Fn('Squeeze', lambda x: jnp.squeeze(x, (1, 2)), n_out=1)
        ]),
        #                                         # vec_e mask_e tok_e vec_d tok_d

        # Encode.
        encoder,  # vec_e mask_e tok_e vec_d tok_d

        # Concat encoder and decoder, given encoder mask.
        encdec_layers,

        # Run decoder blocks.
        _ReversibleSerialForget(decoder_blocks, d_model2, n_layers_forget,
                                forget_dense),  # vec_ed1 vec_ed2 tok_e tok_d
        tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0),  # vec_ed tok_e tok_d
        tl.LayerNorm(),  # vec_ed tok_e tok_d

        # Separate out the encoder part from the concatenated vector.
        tl.Select([0, 1, 2, 2]),  # vec_ed tok_e tok_d tok_d
        t2.StripFromConcatenateWithPadding(mode=mode),  # vec_d tok_d

        # Map to output vocab.
        dense_loss_layer,  # vec_d tok_d
    )