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:: python
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.
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.
activation (str, optional): The activation. Default "gelu".
training (bool, optional): A flag indicating whether it is in train phrase or not. Default False.
