o "j@sddlZddlmZmZddlmZddlmZmZddlm Z ddl m Z ddl m Z mZgZdd Z           dddZ      dddZ              dddZ             dddZdS)N)_C_ops _legacy_C_ops)core) check_dtypecheck_variable_and_dtype)default_main_program) LayerHelper)in_dynamic_modein_dynamic_or_pir_modecCs2t|ttfs td|dks|dkrtddS)Nz(dropout_rate argument should be a numberrz,dropout_rate argument should between 0 and 1) isinstancefloatint TypeError ValueError) dropout_raterp/var/www/html/Deteccion_Ine/venv/lib/python3.10/site-packages/paddle/incubate/nn/functional/fused_transformer.py_verify_dropout_rates r?reluh㈵>FTupscale_in_trainc#Csvt| t| d}|dvrtd|dkrdn|}trtjjdkr+tjj}trtj g|dd||||||||d|d| d | d | d | d | d | d|dud|dud|durx|ndd|dur|ndd|d|d|d|R\ }}}}}}}}}}}|S|j }t |dgddt |dgddt |dd|||||||||| | | | | ||| |du|du|dur|nd|dur|nd||\ }}}}}}}}}}}|Std}|j }t |dgddt |dgdd||j }|jddd}|jddd}|j|j dd}|j|j dd}|j|j dd}|j|j dd}|j|j dd}|j|j dd} |j|j dd}!|j|j dd}"|dusl|dkrw|jjdkrw|jj}|jd|||||||||d ||||||||| |!|"d | | | || | | |du|du|dur|nd|dur|nd||||dd |S)!a This is a fusion operator to compute feed forward layer in transformer model architecture. This operator only supports running on GPU. The function of the operator is consistent with the following pseudo code: .. code-block:: text >>> residual = x >>> if pre_layer_norm: ... out = layer_norm1(x) ... else: ... out = x >>> out = linear2(dropout1(activation(linear1(src)))) >>> if add_residual: ... out = residual + dropout2(out) ... else: ... out = dropout2(out) >>> if not pre_layer_norm: ... out = layer_norm2(out) Args: x (Tensor): the input tensor could be 3-D tensor, the input data type could be float16, float32 or float64, the shape is`[batch\_size, sequence\_length, d_model]`. linear1_weight (Tensor): The weight of first linear, the data type is same as `x`, the shape is `[d\_model, dim\_feedforward]`. linear2_weight (Tensor): The weight of second linear, the data type is same as `x`, the shape is `[dim\_feedforward, d\_model]`. linear1_bias (Tensor, optional): The bias of first linear, the data type is same as `x`, the shape is `[dim_feedforward]`. Default None. linear2_bias (Tensor, optional): The bias of second linear, the data type is same as `x`, the shape is `[d_model]`. Default None. ln1_scale (Tensor, optional): the weight of first layer_norm, the data type is float32 or float64, the shape is same as `x`. Default None. ln1_bias (Tensor, optional): The bias of first layer_norm, the data type is float32 or float64, the shape is `[d\_model]`. Default None. ln2_scale (Tensor, optional): The weight of second layer_norm, the data type is float32 or float64, the shape is same as `x`. Default None. ln2_bias (Tensor, optional): The bias of second layer_norm, the data type is float32 or float64, the shape is `[d\_model]`. Default None. dropout1_rate (float, optional): The first dropout probability of setting units to zero. Default 0.5. dropout2_rate (float, optional): The second dropout probability of setting units to zero. Default 0.5. activation (str, optional): The activation. Default "relu". ln1_epsilon (float, optional): Small float of first layer_norm added to denominator to avoid dividing by zero. Default is 1e-5. ln2_epsilon (float, optional): Small float of second layer_norm added to denominator to avoid dividing by zero. Default is 1e-5. pre_layer_norm (bool, optional): add layer_norm in the pre-processing stage or post-processing state. training (bool, optional): A flag indicating whether it is in train phrase or not. Default True. mode (str, optional): ['upscale_in_train'(default) | 'downscale_in_infer'] 1. upscale_in_train(default), upscale the output at training time - train: out = input * mask / ( 1.0 - p ) - inference: out = input 2. downscale_in_infer, downscale the output at inference - train: out = input * mask - inference: out = input * (1.0 - p) ring_id (int, optional): For distributed forward in tensor model parallel, only support NCCL. Default is -1, means not using tensor parallel. add_residual (bool, optional): Whether add residual at the end. Default is True. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Returns: Tensor: The output Tensor, the data type and shape is same as `x`. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env:GPU) >>> import paddle >>> paddle.device.set_device('gpu') >>> import paddle.incubate.nn.functional as F >>> x = paddle.randn(shape=(1, 8, 8), dtype="float32") >>> linear1_weight = paddle.randn(shape=(8, 8), dtype="float32") >>> linear2_weight = paddle.randn(shape=(8, 8), dtype="float32") >>> out = F.fused_feedforward(x, linear1_weight, linear2_weight) >>> print(out.shape) [1, 8, 8] Ndownscale_in_inferrBmode argument should be 'downscale_in_infer' or 'upscale_in_train'rdowngrade_in_inferrpre_layer_norm ln1_epsilon ln2_epsilon act_method dropout1_rate dropout2_rateis_testdropout1_fix_seeddropout2_fix_seed dropout1_seed dropout2_seeddropout1_implementationdropout2_implementation add_residualring_idxfloat16float32Zfloat64fused_feedforwarddtypeZuint8T) stop_gradient) XZ Linear1WeightZ Linear1BiasZ Linear2WeightZ Linear2BiasZLn1ScaleZLn1BiasLn2ScaleLn2Bias) OutZ Dropout1MaskZ Dropout2MaskZLn1MeanZ Ln1VarianceLn2Mean Ln2VarianceZ Linear1OutZLn1OutZ Dropout1OutZ Dropout2Out)r"r#r!rrr r$r%r&r'r(r)r*r+r,typeinputsoutputsattrs)rrr paddlestaticr random_seedr rr1r2rrrr"create_variable_for_type_inference main_program append_op)#r-Zlinear1_weightZlinear2_weightZ linear1_biasZ linear2_biasZ ln1_scaleZln1_biasZ ln2_scaleZln2_biasr"r# activationrr rtrainingmoder,r+nameseedout_r2helperZ dropout1_maskZ dropout2_maskZln1_meanZ ln1_varianceZln2_meanZ ln2_varianceZ linear1_outZln1_outZ dropout1_outZ dropout2_outrrrr1$s]       !"#$%&'()T    " ,r1c CsDd} |dvr td|dkrdn|}|dur4t|jdks!Jd|jt|jd|jdks4Jd |durVt|jdksCJd |jt|jd|jdksVJd trtjdkrctj} t|||||d |d |d| d| dud| dur|| ndd|\} } } } } | St d!it } |j }t |dgddt |dgddi}|g|d<|g|d<|dur|g|d<|r|g|d<|r|g|d<| dus| dkr| j jdkr| j j} ||| | du| dur| nd|d}| jtjjjdd}| j|dd}| j|dd}| j|d}| j|d} | jd|||||| d|d | S)"a The fused_bias_dropout_residual_layer_norm operator. The pseudo code is as follows: .. code-block:: text >>> y = layer_norm(residual + dropout(bias + x)) Parameters: x (Tensor): The input tensor. The shape is `[*, embed\_dim]`. residual (Tensor): The residual tensor. The shape is same as x. bias (Tensor, optional): The bias of linear. The shape is `[embed_dim]`. Default None. ln_scale (Tensor, optional): The weight tensor of layernorm. The shape is `[embed_dim]`. Default None. ln_bias (Tensor, optional): The bias tensor of layernorm. The shape is `[embed_dim]`. Default None. dropout_rate (float, optional): The dropout probability used on attention weights to drop some attention targets for the dropout after attention. 0 for no dropout. Default 0.5. ln_epsilon (float, optional): Small float value added to denominator of layer_norm to avoid dividing by zero. Default is 1e-5. training (bool, optional): A flag indicating whether it is in train phrase or not. Default True. mode (str, optional): ['upscale_in_train'(default) | 'downscale_in_infer'] 1. upscale_in_train(default), upscale the output at training time - train: out = input * mask / ( 1.0 - p ) - inference: out = input 2. downscale_in_infer, downscale the output at inference - train: out = input * mask - inference: out = input * (1.0 - p) name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Returns: Tensor, The output Tensor, the data type and shape is same as `x`. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env:GPU) >>> import paddle >>> paddle.device.set_device('gpu') >>> import paddle.incubate.nn.functional as F >>> # input: [batch_size, seq_len, embed_dim] >>> x = paddle.rand(shape=(2, 4, 128), dtype="float32") >>> # residual: [batch_size, seq_len, embed_dim] >>> residual = paddle.rand(shape=(2, 4, 128), dtype="float32") >>> # linear bias: [embed_dim] >>> bias = paddle.rand(shape=[128], dtype="float32") >>> # output: [batch_size, seq_len, embed_dim] >>> output = F.fused_bias_dropout_residual_layer_norm( ... x, residual, bias) >>> print(output.shape) [2, 4, 128] Nrrrrr z.The dims of the shape of ln_scale should be 1.rz4The dim of ln_scale must equal to the last dim of x.z-The dims of the shape of ln_bias should be 1.z3The dim of ln_bias must equal to the last dim of x.r ln_epsilonr$dropout_fix_seed dropout_seeddropout_implementation&fused_bias_dropout_residual_layer_normr-r.r2r4ZResidualZBiasLnScaleLnBias)rMrr$rNrOrPTr2r3r2)BiasDropoutResidualOutDropoutMaskOutLnMean LnVarianceYr:)rQ)rlenshaper rrArrQrlocalsr2rrrCrBrVarDescVarTypeUINT8rD)r-ZresidualZbiasln_scaleln_biasrrMrFrGrHrIrK final_outrLr2r<r>dropout_mask_out ln_mean_outln_variance_outbias_dropout_residual_outrrrrQCsE           rQc1Csdd}|dvr td|dkrdn|}|jdkr td|jdtrtjjd kr2tjj}|spt|jd ks?Jd |jd dksJJd |jd|jd ksXJd|dkro|jd|jd |jdksoJdnR|d ksxJdt|jd ksJd|dkr|jdd|jd ksJd|jd |jd ksJd| durt| jdksJd| jd |jdksJdt rJt j g||||| | | || ||d|d|d|d|d| d|d|d| d |dud!|dud"|dur|nd d#|dur |nd d$|d%|d&|d'|R\}}}}}}}}}}}}}}}}}}}}nFt ||||| | | || |||||||| |du|durh|nd || |du|duru|nd ||||\}}}}}}}}}}}}}}}}}}}}| dur||fS|St dOit}|j}t|d)gd*d+t|d,gd*d(i}|g|d-<|r|g|d.<|r|g|d/<|g|d0<| dur| g|d1<| |d2<|g|d3<| dur| g|d4<|r|g|d5<|r|g|d6<| r | g|d7<|dus|d kr|jjd kr|jj}id|d|d|d| d|d| d |dud!|dud"|durD|nd d#|durN|nd d$|d%|d&|d'|d|d|}|j|d8d9}|j|d8d9} |j|d:}!|j|d:}"|j|d:}#|j|d:}$|j|d:}%|j|d:}&|j|d:}'|jtjjjd8d9}(|j|d:})|j|d:}*|j|d:}+|j|d:},|jtjjjd8d9}-|j|d8d9}.|j|d8d9}/|j|d:}0|j|d:}|j|d:}|jd;|id<|d=| d>|!d?|"d@|#dA|$dB|%dC|&dD|'dE|(dF|)dG|*dH|+dI|,dJ|-dK|.dL|/|0||dM|dN| r0||fS|S)Pa Attention mapps queries and a set of key-value pairs to outputs, and Multi-Head Attention performs multiple parallel attention to jointly attending to information from different representation subspaces. This API only support self_attention. The pseudo code is as follows: .. code-block:: text >>> residual = x >>> if pre_layer_norm: ... out = layer_norm(x) ... else: ... out = x >>> # compute q, k, v >>> out = matmul(out, qkv_weight) + qkv_bias >>> out = transpose(out, perm=[2, 0, 3, 1, 4]) >>> # extract q, k and v from out >>> q = out[0:1,::] * (head_dim ** -0.5) >>> k = out[1:2,::] >>> v = out[2:3,::] >>> out = matmul(q, k, transpose_y=True) >>> out = out + attn_mask >>> out = softmax(out) >>> out = dropout(out) >>> out = matmul(out, v) >>> # combine heads >>> out = transpose(out, perm=[0, 2, 1, 3]) >>> # project to output >>> out = linear(out) >>> if add_residual: ... out = residual + dropout(out) ... else: ... out = dropout(out) >>> if not pre_layer_norm: ... out = layer_norm(out) Parameters: x (Tensor): The input tensor of fused_multi_head_attention. The shape is `[batch\_size, sequence\_len, embed\_dim]`. qkv_weight (Tensor): The qkv weight tensor. If `transpose_qkv_wb` is False, the shape is `[3, num_head, dim_head, dim_embed]`. Otherwise, the shape is `[dim_embed, 3 * dim_embed]`. linear_weight (Tensor): The linear weight tensor. The shape is `[embed_dim, embed_dim]`. pre_layer_norm (bool, optional): whether it is pre_layer_norm (True) or post_layer_norm architecture (False). Default False. pre_ln_scale (Tensor, optional): The weight tensor of pre layernorm. Default None. pre_ln_bias (Tensor, optional): The bias tensor of pre layernorm. Default None. ln_scale (Tensor, optional): The weight tensor of layernorm. Default None. ln_bias (Tensor, optional): The bias tensor of layernorm. Default None. pre_ln_epsilon (float, optional): Small float value added to denominator of the pre layer_norm to avoid dividing by zero. Default is 1e-5. qkv_bias (Tensor, optional): The bias of qkv computation. If `transpose_qkv_wb` is False, the shape is `[3, num_head, dim_head]`. Otherwise, the shape is `[3 * dim_embed]`. Default None. linear_bias (Tensor, optional): The bias of linear. The shape is `[embed_dim]`. Default None. cache_kv (Tensor, optional): For generation model, cache structure. The shape is `[2, bsz, num_head, seq_len, head_dim]`. Default None. attn_mask (Tensor, optional): A tensor used in multi-head attention to prevents attention to some unwanted positions, usually the paddings or the subsequent positions. It is a tensor with shape broadcasted to `[batch_size, n_head, sequence_length, sequence_length]`. When the data type is bool, the unwanted positions have `False` values and the others have `True` values. When the data type is int, the unwanted positions have 0 values and the others have 1 values. When the data type is float, the unwanted positions have `-INF` values and the others have 0 values. It can be None when nothing wanted or needed to be prevented attention to. Default None. dropout_rate (float, optional): The dropout probability used on attention weights to drop some attention targets for the dropout after attention. 0 for no dropout. Default 0.5. attn_dropout_rate (float, optional): The dropout probability used on attention weights to drop some attention targets for the dropout in attention. 0 for no dropout. Default 0.5. ln_epsilon (float, optional): Small float value added to denominator of layer_norm to avoid dividing by zero. Default is 1e-5. training (bool, optional): A flag indicating whether it is in train phrase or not. Default True. mode (str, optional): ['upscale_in_train'(default) | 'downscale_in_infer'] 1. upscale_in_train(default), upscale the output at training time - train: out = input * mask / ( 1.0 - p ) - inference: out = input 2. downscale_in_infer, downscale the output at inference - train: out = input * mask - inference: out = input * (1.0 - p) ring_id (int, optional): For distributed forward in mp, only support NCCL and forward. Default is -1, means not using mp add_residual (bool, optional): Whether add residual at the end. Default is True. num_heads (int, optional): If enable transpose_qkv_wb, should provide the num_heads. Default is -1, means not transpose qkv wb. transpose_qkv_wb (bool, optional): Whether transpose the qkv_weight and qkv_bias in the op. Only support GPU for now. Default is false, means not transpose qkv wb. name (str, optional): For details, please refer to :ref:`api_guide_Name`. Generally, no setting is required. Default: None. Returns: Tensor: The output Tensor, the data type and shape is same as `x`. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env:GPU) >>> import paddle >>> paddle.device.set_device('gpu') >>> import paddle.incubate.nn.functional as F >>> # input: [batch_size, seq_len, embed_dim] >>> x = paddle.rand(shape=(2, 4, 128), dtype="float32") >>> # qkv_weight: [3, num_head, head_dim, embed_dim] >>> qkv_weight = paddle.rand(shape=(3, 4, 32, 128), dtype="float32") >>> # qkv_bias: [3, num_head, head_dim] >>> qkv_bias = paddle.rand(shape=(3, 4, 32), dtype="float32") >>> # linear_weight: [embed_dim, embed_dim] >>> linear_weight = paddle.rand(shape=(128, 128), dtype="float32") >>> # linear_bias: [embed_dim] >>> linear_bias = paddle.rand(shape=[128], dtype="float32") >>> # self attention mask: [batch_size, num_heads, seq_len, seq_len] >>> attn_mask = paddle.rand(shape=(2, 4, 4, 4), dtype="float32") >>> # output: [batch_size, seq_len, embed_dim] >>> output = F.fused_multi_head_attention( ... x, qkv_weight, linear_weight, False, ... None, None, None, None, 1e-5, qkv_bias, ... linear_bias, None, attn_mask) >>> print(output.shape) [2, 4, 128] Nrrrrz,The rank of the x should be 3, but received .rz0The dims of the shape of qkv_weight should be 4.zEThe shape of qkv_weight should be [3, num_head, head_dim, embed_dim].zOThe 3rd dim of qkv_weight and 2nd dim of x should be the same, i.e., embed_dim.rr z)embed_dim must be divisible by num_heads.zRWhen enable transpose_qkv_wb, the num_heads should be provided and greater than 0.zkWhen enable transpose_qkv_wb, the dims of the shape of qkv_weight should be 2 when enable transpose_qkv_wb.zxWhen enable transpose_qkv_wb, the shape of qkv_weight should be [embed_dim, 3 * embed_dim] when enable transpose_qkv_wb.zmWhen enable transpose_qkv_wb, the 1st dim of qkv_weight and 2nd dim of x should be the same, i.e., embed_dim.zLWhen enable transpose_qkv_wb, the dims of the shape of qkv_bias should be 1.ztWhen enable transpose_qkv_wb, the 1st dim of qkv_bias and 2nd dim of qkv_weight should be the same, i.e., embed_dim. num_headstranspose_qkv_wbrepsilonrattn_dropout_raterMr$Zattn_dropout_fix_seedrNZattn_dropout_seedrOZattn_dropout_implementationrPr+r,fused_multi_head_attentionr-r.Zfused_multihead_attentionr2r4rRrSQKVWQKVBiasSrcMask OutLinearW OutLinearBiasr5r6CacheKVTrTrUfused_attentionrXrYZLnOutZQKVOutZ QKVBiasOutZ TransposeOut2ZQKOutZQKTVOutZ SoftmaxOutZAttnDropoutMaskOutZAttnDropoutOutZ SrcMaskOutZFMHAOutZ OutLinearOutrWr8r9)rVrZ CacheKVOutr:)rp)rndimr r?r@rrAr[r\r rrwrrr]r2rrrCrBrr^r_r`rD)1r-Z qkv_weightZ linear_weightrZ pre_ln_scaleZ pre_ln_biasrarbZpre_ln_epsilonZqkv_biasZ linear_biasZcache_kv attn_maskrrorMrFrGr,r+rlrmrHrIrK cache_kv_outrcrLr2r<r>Zpre_ln_mean_outZpre_ln_variance_outZ pre_ln_outZqkv_outZ qkv_bias_outZ transpose_outZqk_outZqktv_outZ softmax_outZattn_dropout_mask_outZattn_dropout_outZ attn_mask_outZfmha_outZout_linear_outrdrerfrgrrrrps          !"#$%&'()*+D             "                        rpgeluc$ Csv|dvrtd|dkrdn|}trrtjg|||||||||||||||| | | | |d| d|d|d|d | d |d |d |d |R\}}|durp||fS|Std-it}|j} t|dddgdt| dddgdi}!|g|!d<||!d<||!d<||!d<|dur||!d<|durt |t |ksJ||!d<|dur||!d<|dur||!d<|dkr||!d<||!d<||!d<||!d <|dur||!d!<||!d"<||!d#<| |!d$<| dur| |!d%<| |!d&<| dur | |!d'<| |||| ||||d( }"i}#|j | d)}||#d*<|r)||#d+<|j d|!|#|"d,|r9||fS|S).a" This is a fusion operator to compute multi transformer layers in transformer model architecture. This operator only supports running on GPU. The function of the transformer layer is consistent with the following pseudo code: .. code-block:: text >>> if pre_layer_norm: ... out = layer_norm(x) ... out = qkv_linear(out) + qkv_bias ... else: ... out = qkv_linear(x) + qkv_bias >>> out = transpose(out, perm=[2, 0, 3, 1, 4]) >>> # extract q, k and v from out. >>> q = out[0:1, ::] >>> k = out[1:2, ::] >>> v = out[2:3, ::] >>> out = q * k^t >>> out = attn_mask + out >>> out = softmax(out) >>> out = dropout(out) >>> out = out * v >>> out = transpose(out, perm=[0, 2, 1, 3]) >>> out = linear(out) >>> if pre_layer_norm: ... out = x + dropout(out + bias) ... else: ... out = layer_norm(x + dropout(out + bias)) >>> residual = out; >>> if pre_layer_norm: ... out = ffn_layer_norm(out) >>> out = ffn1_linear(out) >>> out = dropout(activation(out + ffn1_bias)) >>> out = ffn2_linear(out) >>> out = residual + dropout(out + ffn2_bias) >>> if not pre_layer_norm: ... out = ffn_layer_norm(out) Args: x (Tensor): the input tensor could be 3-D tensor, the input data type could be float16 or float32, the shape is `[batch\_size, sequence\_length, d\_model]`. ln_scales (list(Tensor)|tuple(Tensor)): The weight tensors of attention layer_norm, the shape is `[d\_model]`. ln_biases (list(Tensor)|tuple(Tensor)): The bias tensors of attention layer_norm. the shape is `[d\_model]`. qkv_weights (list(Tensor)|tuple(Tensor)): The weight tensors of attention qkv computation. The shape is `[3, num\_head, dim\_head, d\_model]`. qkv_biases (list(Tensor)|tuple(Tensor)|None): The bias tensors of attention qkv computation. The shape is `[3, num\_head, dim\_head]`. linear_weights (list(Tensor)|tuple(Tensor)): The weight tensors of attention linear. The shape is `[num\_head * dim\_head, d\_model]`. linear_biases (list(Tensor)|tuple(Tensor)|None): The bias tensors of attention linear. The shape is `[d\_model]`. ffn_ln_scales (list(Tensor)|tuple(Tensor)): The weight tensors of feedforward layer_norm, the shape is `[d\_model]` ffn_ln_biases (list(Tensor)|tuple(Tensor)): The bias tensors of feedforward layer_norm, the shape is `[d\_model]` ffn1_weights (list(Tensor)|tuple(Tensor)): The weight tensors of feedforward first linear, the shape is `[d\_model, dim\_feedforward]`. ffn1_biases (list(Tensor)|tuple(Tensor)|None): The bias tensors of feedforward first linear, the shape is `[dim\_feedforward]`. ffn2_weights (list(Tensor)|tuple(Tensor)): The weight tensors of feedforward second linear, the shape is `[dim\_feedforward, d\_model]`. ffn2_biases (list(Tensor)|tuple(Tensor)|None): The bias tensors of feedforward second linear, the shape is `[d_model]`. pre_layer_norm (bool, optional): whether it is pre_layer_norm(True) or post_layer_norm(False). Default True. epsilon (float, optional): Small float value added to denominator of the layer_norm to avoid dividing by zero. Default is 1e-5. cache_kvs (list(Tensor)|tuple(Tensor), optional): The cache structure tensors for the generation model. The shape is `[2, bsz, num\_head, max\_seq\_len, head\_dim]`. Default None. pre_caches (list(Tensor)|tuple(Tensor), optional): The prefix caches for the generation model. The shape is `[2, bsz, num\_head, cache\_len, head\_dim]`. Default None. seq_lens (Tensor optional): The sequence lengths of this batch. The shape is `[bsz]`. Default None. rotary_embs (Tensor optional): The RoPE embs for rotary computation. The shape is `[2, bsz, 1, seq\_len, head\_dim]`. Default None. time_step (Tensor, optional): The time step tensor for the generation model. Which used in decode stage, to represent the time step, that is, the real seq_len of CacheKV. The shape is `[1]`, must be in CPUPlace. Default None. attn_mask (Tensor, optional): A tensor used in multi-head attention to prevents attention to some unwanted positions, usually the paddings or the subsequent positions. It is a tensor with shape `[batch_size, 1, sequence_length, sequence_length]`. Default None. dropout_rate (float, optional): The dropout probability of setting units to zero. Default 0.0. rotary_emb_dims (int, optional): The rotary_emb_dims of rotary computation, and it is 0 when rotary_embs is None, 1 when rotary_embs is not None and pos_extra_ids is None, 2 when rotary_embs and pos_extra_ids are both not None. Default 0. activation (str, optional): The activation. Default "gelu". training (bool, optional): A flag indicating whether it is in train phrase or not. Default False. mode (str, optional): ['upscale_in_train'(default) | 'downscale_in_infer'] 1. upscale_in_train(default), upscale the output at training time - train: out = input * mask / ( 1.0 - p ) - inference: out = input 2. downscale_in_infer, downscale the output at inference - train: out = input * mask - inference: out = input * (1.0 - p) trans_qkvw (bool, optional): Whether to transpose for weights of qkv. If true, the shape eights of qkv should be [3, num_head, dim_head, dim_embed]. Otherwise the shape of weights of qkv should be [dim_embed, 3, num_head, dim_head]. Default True. ring_id (int, optional): For distributed forward in tensor model parallel, only support NCCL. Default is -1, means not using mp. name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`. Returns: Tensor|tuple: If `cache_kvs` is None, return a tensor that has the same shape and data type with `x`, representing the output of Transformer layers. If `cache_kvs` is not None, return the tuple (output, cache_kvs), which output is the output of Transformer layers, cache_kvs is inplace with input `cache_kvs`. Examples: .. code-block:: python >>> # doctest: +REQUIRES(env:GPU) >>> import paddle >>> paddle.device.set_device('gpu') >>> import paddle.incubate.nn.functional as F >>> # input: [batch_size, seq_len, embed_dim] >>> x = paddle.rand(shape=(2, 4, 128), dtype="float32") >>> # ln_scale: [embed_dim], ln_bias: [embed_dim] >>> ln_scale = paddle.rand(shape=(128,), dtype="float32") >>> ln_bias = paddle.rand(shape=(128,), dtype="float32") >>> # qkv_weight: [3, num_head, head_dim, embed_dim], qkv_bias: [3, num_head, head_dim] >>> qkv_weight = paddle.rand(shape=(3, 4, 32, 128), dtype="float32") >>> qkv_bias = paddle.rand(shape=(3, 4, 32), dtype="float32") >>> # linear_weight: [embed_dim, embed_dim], linear_bias: [embed_dim] >>> linear_weight = paddle.rand(shape=(128, 128), dtype="float32") >>> linear_bias = paddle.rand(shape=(128,), dtype="float32") >>> # ffn_ln_scale: [embed_dim], ffn_ln_bias: [embed_dim] >>> ffn_ln_scale = paddle.rand(shape=(128,), dtype="float32") >>> ffn_ln_bias = paddle.rand(shape=(128,), dtype="float32") >>> # ffn1_weight: [embed_dim, 4*embed_dim], ffn1_bias: [4*embed_dim] >>> ffn1_weight = paddle.rand(shape=(128, 4*128), dtype="float32") >>> ffn1_bias = paddle.rand(shape=(4*128,), dtype="float32") >>> # ffn2_weight: [4*embed_dim, embed_dim], ffn2_bias: [embed_dim] >>> ffn2_weight = paddle.rand(shape=(4*128, 128), dtype="float32") >>> ffn2_bias = paddle.rand(shape=(128,), dtype="float32") >>> # self attention mask: [batch_size, 1, seq_len, seq_len] >>> attn_mask = paddle.rand(shape=(2, 1, 4, 4), dtype="float32") >>> # output: [batch_size, seq_len, embed_dim] >>> output = F.fused_multi_transformer( ... x, [ln_scale], [ln_bias], [qkv_weight], [qkv_bias], ... [linear_weight], [linear_bias], [ffn_ln_scale], [ffn_ln_bias], ... [ffn1_weight], [ffn1_bias], [ffn2_weight], [ffn2_bias], ... attn_mask=attn_mask) >>> print(output.shape) [2, 4, 128] rrrrrrnrrotary_emb_dimsr$rPr! trans_qkvwr,Nfused_multi_transformerr-r/r0r2r4rRrSrqrrrvZTimeStepZ PreCachesrZ RotaryPosEmbZ SeqLengthsrsrtruZ FFNLnScaleZ FFNLnBiasZ FFN1WeightZFFN1BiasZ FFN2WeightZFFN2Bias) rrnrr~r$rPr!rr,rUr7rxr:)r) rr rrrr]r2rrr[rBrD)$r-Z ln_scalesZ ln_biasesZ qkv_weightsZ qkv_biasesZlinear_weightsZ linear_biasesZ ffn_ln_scalesZ ffn_ln_biasesZ ffn1_weightsZ ffn1_biasesZ ffn2_weightsZ ffn2_biasesrrnZ cache_kvsZ pre_cachesZseq_lensZ rotary_embsZ time_steprzrr~rErFrGrr,rHr{rcrLr2r<r>r=rrrrs&+      !"#$%& (     r)NNNNNNrrrrrFTrrTN)NNNrrTrN)FNNNNrNNNNrrrTrrTrFN)TrNNNNNNr|rr}FrTrN)r?rrZ paddle.baserZpaddle.base.data_feederrrZpaddle.base.frameworkrZpaddle.base.layer_helperrZpaddle.frameworkr r __all__rr1rQrprrrrrs     $ D _