def attention_bias_local_2d_block(mesh,
h_dim,
w_dim,
memory_h_dim,
memory_w_dim,
dtype=tf.int32):
"""Bias for attention for local blocks where attention to right is disallowed.
Create the bias matrix by using two separate masks, one for the memory part
which doesn't overlap with the query and second which interacts with the query
and should be disallowed to look to the right of the current query position.
Args:
mesh: a MeshTensorflow object
h_dim: a mtf.Dimension
w_dim: a mtf.Dimension
memory_h_dim: a mtf.Dimension
memory_w_dim: a mtf.Dimension
dtype: a tf.dtype
Returns:
a mtf.Tensor with shape [block_length, memory_length]
"""
memory_height = mtf.Dimension(memory_h_dim.name, h_dim.size)
memory_width = mtf.Dimension(memory_w_dim.name, w_dim.size)
mask_top_visible = mtf.zeros(mesh, [h_dim, memory_height], dtype=dtype)
mask_left_visible = mtf.zeros(mesh, [w_dim, memory_width], dtype=dtype)
mask_query = mtf.greater(
mtf.range(mesh, memory_height, dtype=tf.int32),
mtf.range(mesh, memory_width, dtype=dtype))
width_mask = mtf.concat([mask_left_visible, mask_query], memory_width.name)
mask = mtf.cast(
mtf.concat([mask_top_visible, width_mask], memory_height.name),
dtype=tf.float32) * -1e9
return mask
def _is_finished(i, unused_alive_seq, alive_log_probs, unused_finished_seq,
finished_scores, finished_in_finished, *unused_states):
"""Checking termination condition.
We terminate when we decoded up to decode_length or the lowest scoring item
in finished has a greater score that the highest prob item in alive divided
by the max length penalty
Args:
i: loop index
alive_log_probs: probabilities of the beams. [batch_size, beam_size]
finished_scores: scores for each of these sequences.
[batch_size, beam_size]
finished_in_finished: finished bools for each of these sequences.
[batch_size, beam_size]
Returns:
Bool.
"""
# TODO(noam): support a different decode length...
# decode_length = mtf.constant(mesh, length_dim.size, dtype=tf.int32)
# del alive_log_probs, finished_scores, finished_in_finished
# return mtf.less(i, length_dim.size)
if not stop_early:
return mtf.less(i, decode_length)
max_length_penalty = mtf.pow(
((5. + mtf.cast(decode_length, finished_scores.dtype)) / 6.),
alpha)
# The best possible score of the most likely alive sequence.
lower_bound_alive_scores = mtf.gather(
alive_log_probs, mtf.constant(mesh, 0, dtype=tf.int32),
beam_dim) / max_length_penalty
# Now to compute the lowest score of a finished sequence in finished
# If the sequence isn't finished, we multiply it's score by 0. since
# scores are all -ve, taking the min will give us the score of the lowest
# finished item.
lowest_score_of_finished_in_finished = mtf.reduce_min(
finished_scores *
mtf.cast(finished_in_finished, finished_scores.dtype),
reduced_dim=beam_dim)
# If none of the sequences have finished, then the min will be 0 and
# we have to replace it by -ve INF if it is. The score of any seq in alive
# will be much higher than -ve INF and the termination condition will not
# be met.
lowest_score_of_finished_in_finished += ((1. - mtf.cast(
mtf.reduce_any(finished_in_finished, reduced_dim=beam_dim),
finished_scores.dtype)) * -INF)
bound_is_met = mtf.reduce_all(
mtf.greater(lowest_score_of_finished_in_finished,
lower_bound_alive_scores))
return mtf.logical_and(mtf.less(i, decode_length),
mtf.logical_not(bound_is_met))
def masked_local_attention_1d(x,
kv_channels,
heads,
window_size=128,
master_dtype=tf.float32,
slice_dtype=tf.float32,
length_per_split=None,
return_kv=None,
params=None,
name=None):
"""Attention to the source position and a neighborhood to the left of it.
Attention for a given query position p can only see memory positions
in the range (p - window_size, p].
Args:
x: a mtf.Tensor with shape batch_dims + [length, io_channels]
kv_channels: a mtf.Dimension (the size of the key and value vectors)
heads: a mtf.Dimension (the number of heads)
window_size: an integer
master_dtype: a tf.dtype (deprecated - use params arg)
slice_dtype: a tf.dtype (deprecated - use params arg)
length_per_split: an optional integer indicating the part of the length
dimension per processor. You can omit if the length dimension is not
split.
return_kv: an optional list onto which to append the computed k and v.
params: an optional quadruple of Tensors (see multihead_attention_params())
name: an optional string.
Returns:
a Tensor with the same shape as x
Raises:
ValueError: if channels or depth don't match.
"""
with tf.variable_scope(
name, default_name="masked_local_attention_1d", values=[x]):
batch_dims = x.shape.dims[:-2]
length, io_channels = x.shape.dims[-2:]
if params is None:
wq, wk, wv, wo = multihead_attention_vars(
x.mesh, heads, io_channels, kv_channels,
master_dtype, slice_dtype, x.dtype)
else:
wq, wk, wv, wo = params
# Get query q, keys k and values v.
qkv_shape = mtf.Shape(batch_dims + [heads, length, kv_channels])
q = mtf.einsum([x, wq], qkv_shape)
k = mtf.einsum([x, wk], qkv_shape)
v = mtf.einsum([x, wv], qkv_shape)
if return_kv is not None:
return_kv.extend([k, v])
# Choose a suitable block size.
# We choose the greatest divisor of length_per_split less than or equal
# to max(window_size, 128)
if length_per_split is None:
length_per_split = length.size
block_length = max(window_size, 128)
while length_per_split % block_length != 0:
block_length -= 1
query_block_length = mtf.Dimension("query_block_length", block_length)
memory_block_length = mtf.Dimension("memory_block_length", block_length)
# The num_blocks dimension gets the same name as the length dimension,
# so it will be split in the same way.
num_blocks = mtf.Dimension(length.name, length.size // block_length)
q_shape = batch_dims + [heads, num_blocks, query_block_length, kv_channels]
kv_shape = batch_dims + [
heads, num_blocks, memory_block_length, kv_channels]
q = mtf.reshape(q, q_shape)
k = mtf.reshape(k, kv_shape)
v = mtf.reshape(v, kv_shape)
# augment the keys and values for each block with keys and values for
# the previous window_size timesteps.
k = mtf.left_halo_exchange(k, num_blocks, memory_block_length, window_size)
v = mtf.left_halo_exchange(v, num_blocks, memory_block_length, window_size)
padded_memory_block_length = mtf.Dimension(
"memory_block_length", window_size + block_length)
mpos = mtf.range(x.mesh, padded_memory_block_length, tf.float32)
qpos = mtf.range(x.mesh, query_block_length, tf.float32) + window_size
# prevent looking forward
mask = mtf.cast(mtf.greater(mpos, qpos), x.dtype) * -1e9
# prevent looking >=block_length timesteps backward
mask += mtf.cast(mtf.less_equal(mpos, qpos - block_length), x.dtype) * -1e9
# Note: The first window_size-1 positions can see back into pre-time
# where all the keys and values are zero. We could mask this out, but we
# don't.
o = dot_product_attention(q, k, v, mask=mask)
o = mtf.reshape(o, batch_dims + [heads, length, kv_channels])
return mtf.einsum([o, wo], mtf.Shape(batch_dims + [length, io_channels]))
def multihead_self_attention_incremental(query_antecedent,
prev_k,
prev_v,
step_num,
master_dtype,
slice_dtype,
name="multihead_attention"):
"""Incremental self-attention (one decode step).
In order to use only one variable containing the four weight matrices
packed together, we insist that the query and memory antecedents have the
same dimensionality (io_channels) and that the keys and values have the
same dimensionality (kv_channels).
Args:
query_antecedent: a mtf.Tensor with shape [batch..., io_channels]
prev_k: mtf.Tensor with shape [batch..., heads, memory_length, kv_channels]
prev_v: mtf.Tensor with shape [batch..., heads, memory_length, kv_channels]
step_num: mtf Scalar with dtype tf.int32
master_dtype: a tf.dtype
slice_dtype: a tf.dtype
name: an optional string.
Returns:
y: A mtf.Tensor with shape [batch..., io_channels]
new_k: mtf.Tensor with shape [batch..., heads, memory_length, kv_channels]
new_v: mtf.Tensor with shape [batch..., heads, memory_length, kv_channels]
Raises:
ValueError: if the dimensions do not match.
"""
batch_dims = query_antecedent.shape.dims[:-1]
io_channels = query_antecedent.shape.dims[-1]
heads, memory_length, kv_channels = prev_k.shape.dims[-3:]
with tf.variable_scope(name, default_name="multihead_attention"):
wq, wk, wv, wo = multihead_attention_vars(
query_antecedent.mesh, heads, io_channels, kv_channels,
master_dtype, slice_dtype, query_antecedent.dtype)
memory_antecedent = query_antecedent
q = mtf.einsum(
[query_antecedent, wq],
mtf.Shape(batch_dims + [heads, kv_channels]))
k = mtf.einsum(
[memory_antecedent, wk],
mtf.Shape(batch_dims + [heads, kv_channels]))
v = mtf.einsum(
[memory_antecedent, wv],
mtf.Shape(batch_dims + [heads, kv_channels]))
k = prev_k + mtf.multiply(
k, mtf.one_hot(step_num, memory_length, dtype=prev_k.dtype),
output_shape=prev_k.shape)
v = prev_v + mtf.multiply(
v, mtf.one_hot(step_num, memory_length, dtype=prev_v.dtype),
output_shape=prev_v.shape)
mask = mtf.cast(
mtf.greater(mtf.range(
query_antecedent.mesh, memory_length, dtype=tf.int32), step_num),
q.dtype) * -1e9
o = dot_product_attention(q, k, v, mask)
y = mtf.einsum([o, wo], query_antecedent.shape)
return y, k, v
def multihead_self_attention_memory_compressed(x,
mask_right,
compression_factor,
kv_channels,
heads,
dropout=0.0,
dropout_broadcast_dims=None,
master_dtype=tf.float32,
slice_dtype=tf.float32,
name="multihead_attention"):
"""Memory-compressed self-attention.
The memory is first average-pooled (strided) to make it shorter by
a factor of compression_factor.
Args:
x: a mtf.Tensor with shape
[<batch_dims>, query_length, io_channels]
mask_right: a boolean
compression_factor: an integer
kv_channels: a mtf.Dimension (the size of the key and value vectors)
heads: a mtf.Dimension (the number of heads)
dropout: a floating point value
dropout_broadcast_dims: an optional list of mtf.Dimension
master_dtype: a tf.dtype
slice_dtype: a tf.dtype
name: an optional string.
Returns:
A mtf.Tensor with shape [batch, query_length, io_channels]
Raises:
ValueError: if the dimensions do not match.
"""
batch_dims = x.shape.dims[:-2]
length, io_channels = x.shape.dims[-2:]
with tf.variable_scope(name,
default_name="compressed_attention",
values=[x]):
wq, wk, wv, wo = multihead_attention_vars(
x.mesh, heads, io_channels, kv_channels,
master_dtype, slice_dtype, x.dtype)
memory_antecedent = compress_mean(x, length, compression_factor)
memory_antecedent = rename_length_to_memory_length(memory_antecedent)
memory_length = memory_antecedent.shape.dims[-2]
q = mtf.einsum(
[x, wq],
mtf.Shape(batch_dims + [heads, length, kv_channels]))
k = mtf.einsum(
[memory_antecedent, wk],
mtf.Shape(batch_dims + [heads, memory_length, kv_channels]))
v = mtf.einsum(
[memory_antecedent, wv],
mtf.Shape(batch_dims + [heads, memory_length, kv_channels]))
if mask_right:
query_pos = mtf.range(x.mesh, length, dtype=tf.int32)
memory_pos = (
mtf.range(x.mesh, memory_length, dtype=tf.int32) * compression_factor
+ (compression_factor - 1))
mask = mtf.cast(mtf.greater(memory_pos, query_pos), x.dtype) * -1e9
else:
mask = None
o = dot_product_attention(
q, k, v, mask, dropout, dropout_broadcast_dims, extra_logit=0.0)
return mtf.einsum(
[o, wo], mtf.Shape(batch_dims + [length, io_channels]))