def apply(self,
inputs_q,
inputs_kv,
num_heads,
dtype=jnp.float32,
qkv_features=None,
out_features=None,
attention_axis=None,
causal_mask=False,
padding_mask=None,
key_padding_mask=None,
segmentation=None,
key_segmentation=None,
cache=None,
broadcast_dropout=True,
dropout_rng=None,
dropout_rate=0.,
deterministic=False,
precision=None,
kernel_init=nn.linear.default_kernel_init,
bias_init=nn.initializers.zeros,
bias=True,
block_size=50,
max_num_blocks=25,
sort_activation='softmax'):
"""Applies multi-head sinkhorn attention on the input data.
Projects the inputs into multi-headed query, key, and value vectors,
applies dot-product attention and project the results to an output vector.
This can be used for encoder-decoder attention by specifying both `inputs_q`
and `inputs_kv` orfor self-attention by only specifying `inputs_q` and
setting `inputs_kv` to None.
Args:
inputs_q: input queries of shape `[bs, dim1, dim2, ..., dimN, features]`.
inputs_kv: key/values of shape `[bs, dim1, dim2, ..., dimN, features]`
or None for self-attention, inn which case key/values will be derived
from inputs_q.
num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])
should be divisible by the number of heads.
dtype: the dtype of the computation (default: float32)
qkv_features: dimension of the key, query, and value.
out_features: dimension of the last projection
attention_axis: axes over which the attention is applied ( 'None' means
attention over all axes, but batch, heads, and features).
causal_mask: boolean specifying whether to apply a causal mask on the
attention weights. If True, the output at timestep `t` will not depend
on inputs at timesteps strictly greater than `t`.
padding_mask: boolean specifying query tokens that are pad token.
key_padding_mask: boolean specifying key-value tokens that are pad token.
segmentation: segment indices for packed inputs_q data.
key_segmentation: segment indices for packed inputs_kv data.
cache: an instance of `flax.nn.attention.Cache` used for efficient
autoregressive decoding.
broadcast_dropout: bool: use a broadcasted dropout along batch dims.
dropout_rng: JAX PRNGKey: to be used for dropout
dropout_rate: dropout rate
deterministic: bool, deterministic or not (to apply dropout)
precision: numerical precision of the computation see `jax.lax.Precision`
for details.
kernel_init: initializer for the kernel of the Dense layers.
bias_init: initializer for the bias of the Dense layers.
bias: bool: whether pointwise QKVO dense transforms use bias.
block_size: int, block size.
max_num_blocks: int, max num blocks.
sort_activation: str {softmax, sinkhorn, gumbel_sinkhorn}
Returns:
output of shape `[bs, dim1, dim2, ..., dimN, features]`.
"""
assert causal_mask or not cache, (
'Caching is only support for causal attention.')
assert inputs_q.ndim == 3
if inputs_kv is None:
inputs_kv = inputs_q
if attention_axis is None:
attention_axis = tuple(range(1, inputs_q.ndim - 1))
features = out_features or inputs_q.shape[-1]
qkv_features = qkv_features or inputs_q.shape[-1]
assert qkv_features % num_heads == 0, (
'Memory dimension must be divisible by number of heads.')
head_dim = qkv_features // num_heads
dense = nn.DenseGeneral.partial(
axis=-1,
features=(num_heads, head_dim),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision)
# project inputs_q to multi-headed q/k/v
# dimensions are then [bs, dims..., n_heads, n_features_per_head]
qlength = inputs_q.shape[-2]
bs = inputs_q.shape[0]
kvlength = inputs_kv.shape[-2]
query, key, value = (dense(inputs_q, dtype=dtype, name='query'),
dense(inputs_kv, dtype=dtype, name='key'),
dense(inputs_kv, dtype=dtype, name='value'))
if cache:
assert isinstance(cache, Cache), 'cache must be an instance of Cache'
if self.is_initializing():
cache.store(onp.array((key.ndim,) + key.shape[-2:], dtype=onp.int32))
else:
cache_entry = cache.retrieve(None)
expected_shape = list(cache_entry.key.shape[:-2])
for attn_dim in attention_axis:
expected_shape[attn_dim] = 1
expected_shape = tuple(expected_shape) + inputs_q.shape[-1:]
if expected_shape != inputs_q.shape:
raise ValueError('Invalid shape provided, '
'expected shape %s instead got %s.' %
(expected_shape, inputs_q.shape))
if not isinstance(cache_entry, _CacheEntry):
raise ValueError('Cache is not initialized.')
cshape = cache_entry.key.shape
indices = [0] * len(cshape)
i = cache_entry.i
attn_size = onp.prod(onp.take(cshape, attention_axis))
for attn_dim in attention_axis:
attn_size //= cshape[attn_dim]
indices[attn_dim] = i // attn_size
i = i % attn_size
key = lax.dynamic_update_slice(cache_entry.key, key, indices)
value = lax.dynamic_update_slice(cache_entry.value, value, indices)
one = jnp.array(1, jnp.uint32)
cache_entry = cache_entry.replace(i=cache_entry.i + one,
key=key,
value=value)
cache.store(cache_entry)
key_padding_mask = jnp.broadcast_to(
(jnp.arange(cshape[1]) < cache_entry.i), cshape[:2])
key_padding_mask = key_padding_mask.astype(jnp.float32)[..., None]
# block reshape before attention
num_query_blocks = qlength // block_size
num_kv_blocks = kvlength // block_size
block_query = jnp.reshape(
query, (bs, block_size, num_query_blocks, num_heads, head_dim))
block_key = jnp.reshape(
key, (bs, block_size, num_kv_blocks, num_heads, head_dim))
block_value = jnp.reshape(
value, (bs, block_size, num_kv_blocks, num_heads, head_dim))
if causal_mask:
# causal masking needs to not have blocks with mixed information.
sum_key = jnp.cumsum(block_key, axis=1)
sum_key = sum_key[:, 0, :, :, :] # take first item
else:
sum_key = jnp.sum(block_key, axis=1)
# sort net on head_dim dimensions
sort_out = nn.DenseGeneral(sum_key, axis=-1,
features=(max_num_blocks),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision)
# (bs x num_key_blocks x num_heads x num_key_blocks
sort_out = sort_out[:, :, :, :num_query_blocks]
# simple softmax sorting first.
if sort_activation == 'sinkhorn':
permutation = sinkhorn_operator(
jnp.reshape(sort_out, (-1, num_kv_blocks, num_query_blocks)),
causal=causal_mask)
permutation = jnp.reshape(permutation, (-1, num_kv_blocks, num_heads,
num_query_blocks))
else:
if causal_mask:
block_mask = _make_causal_mask(key, attention_axis)
sort_out += block_mask
permutation = jax.nn.softmax(sort_out, axis=-1)
sorted_key = jnp.einsum('bskhd,bnhl->bsnhd', block_key, permutation)
sorted_value = jnp.einsum('bskhd,bnhl->bsnhd', block_value, permutation)
# create attention masks
mask_components = []
sorted_mask_components = []
if causal_mask:
# TODO(yitay): Test this causal masking.
if cache and not self.is_initializing():
bias_pre_shape = (1,) * (key.ndim - 1)
attn_shape = tuple(onp.take(key.shape, attention_axis))
attn_size = onp.prod(attn_shape)
ii = jnp.arange(attn_size, dtype=jnp.uint32)
mask = ii < cache_entry.i
mask_components.append(mask.reshape(bias_pre_shape + attn_shape))
else:
mask_components.append(_make_causal_mask(key, attention_axis))
if padding_mask is not None:
# divide padding mask into block
padding_mask = jnp.reshape(padding_mask,
(bs * num_query_blocks, block_size, 1))
if key_padding_mask is None:
key_padding_mask = padding_mask
padding_mask = make_padding_mask(
padding_mask_query=padding_mask,
padding_mask_key=key_padding_mask,
query_shape=(bs * num_query_blocks, block_size, num_heads, head_dim),
key_shape=(bs * num_kv_blocks, block_size, num_heads, head_dim),
attention_axis=attention_axis)
padding_mask = jnp.reshape(padding_mask,
(bs, num_query_blocks, block_size, block_size))
mask_components.append(padding_mask)
sorted_padding_mask = jnp.einsum('bksj,bnhl->bnsj', padding_mask,
permutation)
sorted_mask_components.append(sorted_padding_mask)
if segmentation is not None:
if key_segmentation is None:
key_segmentation = segmentation
segmentation_mask = make_padding_mask(
padding_mask_query=segmentation,
padding_mask_key=key_segmentation,
query_shape=(bs * num_query_blocks, block_size, num_heads, head_dim),
key_shape=(bs * num_kv_blocks, block_size, num_heads, head_dim),
attention_axis=attention_axis,
segmentation_mask=True)
segmentation_mask = jnp.reshape(segmentation_mask,
(bs, num_query_blocks, block_size,
block_size))
mask_components.append(segmentation_mask)
sorted_segmentation_mask = jnp.einsum('bksj,bnhl->bnsj',
segmentation_mask,
permutation)
sorted_mask_components.append(sorted_segmentation_mask)
if mask_components:
attention_mask = mask_components[0]
for component in mask_components[1:]:
attention_mask = jnp.logical_and(attention_mask, component)
# attention mask in the form of attention bias
attention_bias = lax.select(
attention_mask > 0, jnp.full(attention_mask.shape, 0.).astype(dtype),
jnp.full(attention_mask.shape, -1e10).astype(dtype))
else:
attention_bias = None
if sorted_mask_components:
attention_mask = sorted_mask_components[0]
for component in sorted_mask_components[1:]:
attention_mask = jnp.logical_and(attention_mask, component)
# attention mask in the form of attention bias
sorted_attention_bias = lax.select(
attention_mask > 0, jnp.full(attention_mask.shape, 0.).astype(dtype),
jnp.full(attention_mask.shape, -1e10).astype(dtype))
else:
sorted_attention_bias = None
# apply attention
x = local_dot_product_attention(
block_query,
block_key,
block_value,
dtype=dtype,
axis=attention_axis,
bias=attention_bias,
precision=precision,
dropout_rng=dropout_rng,
dropout_rate=dropout_rate,
broadcast_dropout=broadcast_dropout,
deterministic=deterministic)
sorted_x = local_dot_product_attention(
block_query,
sorted_key,
sorted_value,
dtype=dtype,
axis=attention_axis,
bias=sorted_attention_bias,
precision=precision,
dropout_rng=dropout_rng,
dropout_rate=dropout_rate,
broadcast_dropout=broadcast_dropout,
deterministic=deterministic)
x = x + sorted_x
x = jnp.reshape(x, (bs, qlength, num_heads, head_dim))
# back to the original inputs dimensions
out = nn.DenseGeneral(
x,
features=features,
axis=(-2, -1),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
dtype=dtype,
precision=precision,
name='out')
return out
def apply(self,
inputs_q,
inputs_kv,
num_heads,
dtype=jnp.float32,
qkv_features=None,
out_features=None,
attention_axis=None,
causal_mask=False,
padding_mask=None,
key_padding_mask=None,
segmentation=None,
key_segmentation=None,
cache=None,
broadcast_dropout=True,
dropout_rng=None,
dropout_rate=0.,
deterministic=False,
precision=None,
kernel_init=nn.linear.default_kernel_init,
bias_init=nn.initializers.zeros,
bias=True,
num_partitions=2):
"""Applies multi-head dot product attention on the input data.
Projects the inputs into multi-headed query, key, and value vectors,
applies dot-product attention and project the results to an output vector.
This can be used for encoder-decoder attention by specifying both `inputs_q`
and `inputs_kv` orfor self-attention by only specifying `inputs_q` and
setting `inputs_kv` to None.
Args:
inputs_q: input queries of shape `[bs, dim1, dim2, ..., dimN, features]`.
inputs_kv: key/values of shape `[bs, dim1, dim2, ..., dimN, features]`
or None for self-attention, inn which case key/values will be derived
from inputs_q.
num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])
should be divisible by the number of heads.
dtype: the dtype of the computation (default: float32)
qkv_features: dimension of the key, query, and value.
out_features: dimension of the last projection
attention_axis: axes over which the attention is applied ( 'None' means
attention over all axes, but batch, heads, and features).
causal_mask: boolean specifying whether to apply a causal mask on the
attention weights. If True, the output at timestep `t` will not depend
on inputs at timesteps strictly greater than `t`.
padding_mask: boolean specifying query tokens that are pad token.
key_padding_mask: boolean specifying key-value tokens that are pad token.
segmentation: segment indices for packed inputs_q data.
key_segmentation: segment indices for packed inputs_kv data.
cache: an instance of `flax.nn.attention.Cache` used for efficient
autoregressive decoding.
broadcast_dropout: bool: use a broadcasted dropout along batch dims.
dropout_rng: JAX PRNGKey: to be used for dropout
dropout_rate: dropout rate
deterministic: bool, deterministic or not (to apply dropout)
precision: numerical precision of the computation see `jax.lax.Precision`
for details.
kernel_init: initializer for the kernel of the Dense layers.
bias_init: initializer for the bias of the Dense layers.
bias: bool: whether pointwise QKVO dense transforms use bias.
num_partitions: number of ways to partition (i.e. how many devices to run
across).
Returns:
output of shape `[bs, dim1, dim2, ..., dimN, features]`.
"""
assert causal_mask or not cache, (
'Caching is only support for causal attention.')
if inputs_kv is None:
inputs_kv = inputs_q
if attention_axis is None:
attention_axis = tuple(range(1, inputs_q.ndim - 1))
features = out_features or inputs_q.shape[-1]
qkv_features = qkv_features or inputs_q.shape[-1]
assert qkv_features % num_heads == 0, (
'Memory dimension must be divisible by number of heads.')
head_dim = qkv_features // num_heads
dense = nn.DenseGeneral.partial(axis=-1,
features=(num_heads, head_dim),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision)
# project inputs_q to multi-headed q/k/v
# dimensions are then [bs, dims..., n_heads, n_features_per_head]
query, key, value = (dense(inputs_q, dtype=dtype, name='query'),
dense(inputs_kv, dtype=dtype, name='key'),
dense(inputs_kv, dtype=dtype, name='value'))
if num_partitions > 1:
partitions = P(1, 1, num_partitions, 1)
query = with_sharding_constraint(query, partitions)
key = with_sharding_constraint(key, partitions)
value = with_sharding_constraint(value, partitions)
if cache:
assert isinstance(cache,
Cache), 'cache must be an instance of Cache'
if self.is_initializing():
cache.store(lambda: (key.ndim, key.shape[-2:]))
else:
cache_entry = cache.retrieve(None)
expected_shape = list(cache_entry.key.shape[:-2])
for attn_dim in attention_axis:
expected_shape[attn_dim] = 1
expected_shape = tuple(expected_shape) + inputs_q.shape[-1:]
if expected_shape != inputs_q.shape:
raise ValueError('Invalid shape provided, '
'expected shape %s instead got %s.' %
(expected_shape, inputs_q.shape))
if not isinstance(cache_entry, _CacheEntry):
raise ValueError('Cache is not initialized.')
cshape = cache_entry.key.shape
i = cache_entry.i
one_hot_indices = jax.nn.one_hot(i, cshape[3],
dtype=key.dtype).reshape(
(1, 1, 1, cshape[3]))
key = key.transpose((0, 2, 3, 1))
key = cache_entry.key + key * one_hot_indices
value = value.transpose((0, 2, 3, 1))
value = cache_entry.value + value * one_hot_indices
one = jnp.array(1, jnp.uint32)
cache_entry = cache_entry.replace(i=cache_entry.i + one,
key=key,
value=value)
cache.store(cache_entry)
key = key.transpose((0, 3, 1, 2))
value = value.transpose((0, 3, 1, 2))
cshape = (cshape[0], cshape[3], cshape[1], cshape[2])
# TODO(levskaya): verify this is still needed in translation decoding.
key_padding_mask = jnp.broadcast_to(
(jnp.arange(cshape[1]) < cache_entry.i), cshape[:2])
key_padding_mask = key_padding_mask.astype(jnp.float32)[...,
None]
# create attention masks
mask_components = []
if causal_mask:
if cache and not self.is_initializing():
bias_pre_shape = (1, ) * (key.ndim - 1)
attn_shape = tuple(np.take(key.shape, attention_axis))
attn_size = np.prod(attn_shape)
ii = jnp.arange(attn_size, dtype=jnp.uint32)
mask = ii < cache_entry.i
mask_components.append(
mask.reshape(bias_pre_shape + attn_shape))
else:
mask_components.append(_make_causal_mask(key, attention_axis))
if padding_mask is not None:
if key_padding_mask is None:
key_padding_mask = padding_mask
padding_mask = make_padding_mask(padding_mask_query=padding_mask,
padding_mask_key=key_padding_mask,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis)
mask_components.append(padding_mask)
if segmentation is not None:
if key_segmentation is None:
key_segmentation = segmentation
segmentation_mask = make_padding_mask(
padding_mask_query=segmentation,
padding_mask_key=key_segmentation,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis,
segmentation_mask=True)
mask_components.append(segmentation_mask)
if mask_components:
attention_mask = mask_components[0]
for component in mask_components[1:]:
attention_mask = jnp.logical_and(attention_mask, component)
# attention mask in the form of attention bias
attention_bias = lax.select(
attention_mask > 0,
jnp.full(attention_mask.shape, 0.).astype(dtype),
jnp.full(attention_mask.shape, -1e10).astype(dtype))
else:
attention_bias = None
# apply attention
x = dot_product_attention(query,
key,
value,
dtype=dtype,
axis=attention_axis,
bias=attention_bias,
precision=precision,
dropout_rng=dropout_rng,
dropout_rate=dropout_rate,
broadcast_dropout=broadcast_dropout,
deterministic=deterministic)
# back to the original inputs dimensions
out = nn.DenseGeneral(x,
features=features,
axis=(-2, -1),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
dtype=dtype,
precision=precision,
name='out')
if num_partitions > 1:
x = with_sharding_constraint(x, None)
return out
def __call__(self,
inputs_q,
inputs_kv,
*,
padding_mask,
key_padding_mask,
segmentation=None,
key_segmentation=None):
"""Applies multi-head dot product attention on the input data.
If weight_prec is not None, scales and quantizes weights to signed int with
weight_prec bits.
Projects the inputs into multi-headed query, key, and value vectors,
applies dot-product attention and project the results to an output vector.
This can be used for encoder-decoder attention by specifying both `inputs_q`
and `inputs_kv` or for self-attention by only specifying `inputs_q` and
setting `inputs_kv` to None.
Args:
inputs_q: input queries of shape `[bs, dim1, dim2, ..., dimN, features]`.
inputs_kv: key/values of shape `[bs, dim1, dim2, ..., dimN, features]` or
None for self-attention, inn which case key/values will be derived from
inputs_q.
padding_mask: boolean tensor specifying query tokens that are pad token.
key_padding_mask: boolean tensor specifying key-value tokens that are pad
token.
segmentation: segment indices for packed inputs_q data.
key_segmentation: segment indices for packed inputs_kv data.
Returns:
output of shape `[bs, dim1, dim2, ..., dimN, features]`.
"""
batch_size, query_sequence_length, channel_size = inputs_q.shape
hparams = self.hparams
if inputs_kv is None:
inputs_kv = inputs_q
key_sequence_length = inputs_q.shape[1]
else:
key_sequence_length = inputs_kv.shape[1]
shape_utils.assert_shapes_equal(
inputs_kv.shape,
(batch_size, key_sequence_length, channel_size))
jax_precision = jax.lax.Precision.DEFAULT
if padding_mask is not None:
shape_utils.assert_shapes_equal(
padding_mask.shape, (batch_size, query_sequence_length, 1))
if key_padding_mask is None:
key_padding_mask = padding_mask
else:
shape_utils.assert_shapes_equal(
key_padding_mask.shape, (batch_size, key_sequence_length, 1))
attention_axis = self.attention_axis
if attention_axis is None:
attention_axis = tuple(range(1, inputs_q.ndim - 1))
qkv_features = self.qkv_features
qkv_features = qkv_features or inputs_q.shape[-1]
num_heads = self.num_heads
assert qkv_features % num_heads == 0, (
'Memory dimension must be divisible by number of heads.')
head_dim = qkv_features // num_heads
paxis_name = self.paxis_name
train = self.train
kernel_init = self.kernel_init
bias_init = self.bias_init
use_bias = self.use_bias
dtype = self.dtype
def multi_batch_dense_aqt(inputs, *, name, padding_mask):
batch_size, sequence_length, channel_size = inputs.shape
inputs = inputs.reshape(batch_size * sequence_length, channel_size)
if padding_mask is not None:
padding_mask = padding_mask.reshape(
batch_size * sequence_length, 1)
out = flax_layers.DenseAqt(name=name,
features=num_heads * head_dim,
paxis_name=paxis_name,
train=train,
quant_context=self.quant_context,
hparams=hparams.dense_kqv,
kernel_init=kernel_init,
bias_init=bias_init,
use_bias=use_bias,
dtype=dtype)(inputs,
padding_mask=padding_mask)
return out.reshape(batch_size, sequence_length, num_heads,
head_dim)
# project inputs_q to multi-headed q/k/v
# dimensions are then [bs, sequence_length, n_heads, n_features_per_head]
query = multi_batch_dense_aqt(inputs_q,
name='query',
padding_mask=padding_mask)
key = multi_batch_dense_aqt(inputs_kv,
name='key',
padding_mask=key_padding_mask)
value = multi_batch_dense_aqt(inputs_kv,
name='value',
padding_mask=key_padding_mask)
is_cache_initialized = False
if self.decode:
is_cache_initialized = self.has_variable('cache', 'cached_key')
cached_key = self.variable('cache', 'cached_key', jnp.zeros,
key.shape, key.dtype)
cached_value = self.variable('cache', 'cached_value', jnp.zeros,
value.shape, value.dtype)
cache_index = self.variable('cache', 'cache_index',
lambda: jnp.array(0, dtype=jnp.int32))
if is_cache_initialized:
expected_shape = list(cached_key.value.shape[:-2])
for attn_dim in attention_axis:
expected_shape[attn_dim] = 1
expected_shape = tuple(expected_shape) + inputs_q.shape[-1:]
if expected_shape != inputs_q.shape:
raise ValueError('Invalid shape provided, '
'expected shape %s instead got %s.' %
(expected_shape, inputs_q.shape))
cshape = cached_key.value.shape
indices = [0] * len(cshape)
i = cache_index.value
attn_size = onp.prod(onp.take(cshape, attention_axis))
*batch_dims, max_length, num_heads, depth_per_head = ( # pylint: disable=unused-variable
cached_key.value.shape)
indices = (0, ) * len(batch_dims) + (i, 0, 0)
key = lax.dynamic_update_slice(cached_key.value, key, indices)
value = lax.dynamic_update_slice(cached_value.value, value,
indices)
one = jnp.array(1, jnp.int32)
cache_index.value = cache_index.value + one
cached_key.value = key
cached_value.value = value
# TODO(levskaya): verify this is still needed in translation decoding.
key_padding_mask = jnp.broadcast_to(
(jnp.arange(max_length) < cache_index.value), cshape[:2])
key_padding_mask = key_padding_mask.astype(
jnp.float32)[Ellipsis, None]
# create attention masks
mask_components = []
if self.causal_mask:
if self.decode and is_cache_initialized:
bias_pre_shape = (1, ) * (key.ndim - 1)
attn_shape = tuple(onp.take(key.shape, attention_axis))
attn_size = onp.prod(attn_shape)
ii = jnp.arange(attn_size, dtype=jnp.int32)
mask = ii < cache_index.value
mask_components.append(
mask.reshape(bias_pre_shape + attn_shape))
else:
mask_components.append(_make_causal_mask(key, attention_axis))
if padding_mask is not None:
if key_padding_mask is None:
key_padding_mask = padding_mask
attn_padding_mask = make_padding_mask(
padding_mask_query=padding_mask,
padding_mask_key=key_padding_mask,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis)
mask_components.append(attn_padding_mask)
if segmentation is not None:
if key_segmentation is None:
key_segmentation = segmentation
segmentation_mask = make_padding_mask(
padding_mask_query=segmentation,
padding_mask_key=key_segmentation,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis,
segmentation_mask=True)
mask_components.append(segmentation_mask)
attention_mask = None
if mask_components:
attention_mask = mask_components[0]
for component in mask_components[1:]:
attention_mask = jnp.logical_and(attention_mask, component)
attention_mask = attention_mask.astype(jnp.bool_)
# attention mask in the form of attention bias
attention_bias = jnp.where(
attention_mask,
jnp.full(attention_mask.shape, 0.).astype(dtype),
jnp.full(attention_mask.shape, -1e10).astype(dtype))
else:
attention_bias = None
# Add an extra dimension to the mask corresponding to the head
# dimension. eg, if inputs_q has shape [batch_size, sequence_length,
# n_features], then padding_mask will have a shape
# [batch_size, sequence_length, 1] and query will have shape
# [batch_size, sequence_length, n_heads, n_features_per_head].
# We create query_padding_mask with shape [batch_size, sequence_length,
# 1, 1] to be broadcast-compatible with 'query'.
if padding_mask is not None:
padding_mask = padding_mask[Ellipsis, None]
shape_utils.assert_shapes_equal(
padding_mask.shape, (batch_size, query_sequence_length, 1, 1))
if key_padding_mask is not None:
key_padding_mask = key_padding_mask[Ellipsis, None]
# During prediction, the key padding mask is only going to be
# broadcast-compatible with the key.
shape_utils.assert_shapes_compatible(
key_padding_mask.shape,
(batch_size, key_sequence_length, 1, 1))
# apply attention
attention_fn = self.attention_fn
dropout_rate = self.dropout_rate
broadcast_dropout = self.broadcast_dropout
deterministic = self.deterministic
if not deterministic and self.dropout_rate > 0.0:
dropout_rng = self.make_rng('dropout')
else:
dropout_rng = None
x = attention_fn( # pylint: disable=redundant-keyword-arg
query=query,
key=key,
value=value,
hparams=hparams.attn_acts,
paxis_name=paxis_name,
train=train,
quant_context=self.quant_context,
dtype=dtype,
axis=attention_axis,
bias=attention_bias,
precision=jax_precision,
dropout_rng=dropout_rng,
dropout_rate=dropout_rate,
broadcast_dropout=broadcast_dropout,
deterministic=deterministic,
query_padding_mask=padding_mask,
key_padding_mask=key_padding_mask,
attn_mask=attention_mask)
shape_utils.assert_shapes_equal(
x.shape, (batch_size, query_sequence_length, num_heads, head_dim))
x = x.reshape(batch_size * query_sequence_length, num_heads * head_dim)
if padding_mask is not None:
padding_mask = padding_mask.reshape(
batch_size * query_sequence_length, 1)
# back to the original inputs dimensions
out = flax_layers.DenseAqt(features=channel_size,
hparams=hparams.dense_out,
quant_context=self.quant_context,
paxis_name=paxis_name,
train=train,
kernel_init=kernel_init,
bias_init=bias_init,
use_bias=use_bias,
dtype=dtype,
name='dense_out')(x,
padding_mask=padding_mask)
shape_utils.assert_shapes_equal(
out.shape, (batch_size * query_sequence_length, channel_size))
out = out.reshape(batch_size, query_sequence_length, channel_size)
return out
def apply(self,
inputs_q,
inputs_kv,
num_heads,
dtype=jnp.float32,
qkv_features=None,
out_features=None,
attention_axis=None,
causal_mask=False,
padding_mask=None,
key_padding_mask=None,
segmentation=None,
key_segmentation=None,
cache=None,
broadcast_dropout=True,
dropout_rng=None,
dropout_rate=0.,
deterministic=False,
precision=None,
kernel_init=nn.linear.default_kernel_init,
bias_init=nn.initializers.zeros,
bias=True,
max_length=512,
ignore_dot_product=True,
synthesizer_mode='factorized_random',
k=32):
"""Applies multi-head synthesizer attention on the input data.
Projects the inputs into multi-headed query, key, and value vectors,
applies dot-product attention and project the results to an output vector.
This can be used for encoder-decoder attention by specifying both `inputs_q`
and `inputs_kv` orfor self-attention by only specifying `inputs_q` and
setting `inputs_kv` to None.
Args:
inputs_q: input queries of shape `[bs, dim1, dim2, ..., dimN, features]`.
inputs_kv: key/values of shape `[bs, dim1, dim2, ..., dimN, features]`
or None for self-attention, inn which case key/values will be derived
from inputs_q.
num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])
should be divisible by the number of heads.
dtype: the dtype of the computation (default: float32)
qkv_features: dimension of the key, query, and value.
out_features: dimension of the last projection
attention_axis: axes over which the attention is applied ( 'None' means
attention over all axes, but batch, heads, and features).
causal_mask: boolean specifying whether to apply a causal mask on the
attention weights. If True, the output at timestep `t` will not depend
on inputs at timesteps strictly greater than `t`.
padding_mask: boolean specifying query tokens that are pad token.
key_padding_mask: boolean specifying key-value tokens that are pad token.
segmentation: segment indices for packed inputs_q data.
key_segmentation: segment indices for packed inputs_kv data.
cache: an instance of `flax.nn.attention.Cache` used for efficient
autoregressive decoding.
broadcast_dropout: bool: use a broadcasted dropout along batch dims.
dropout_rng: JAX PRNGKey: to be used for dropout
dropout_rate: dropout rate
deterministic: bool, deterministic or not (to apply dropout)
precision: numerical precision of the computation see `jax.lax.Precision`
for details.
kernel_init: initializer for the kernel of the Dense layers.
bias_init: initializer for the bias of the Dense layers.
bias: bool: whether pointwise QKVO dense transforms use bias.
max_length: int, the maximum supported sequence length.
ignore_dot_product: bool, to ignore the dot product attention or not.
synthesizer_mode: str support 'dense' and 'random' or 'dense+random'
k: int, low rank factorized attention.
Returns:
output of shape `[bs, dim1, dim2, ..., dimN, features]`.
"""
assert causal_mask or not cache, (
'Caching is only support for causal attention.')
assert inputs_q.ndim == 3
if inputs_kv is None:
inputs_kv = inputs_q
if attention_axis is None:
attention_axis = tuple(range(1, inputs_q.ndim - 1))
features = out_features or inputs_q.shape[-1]
qkv_features = qkv_features or inputs_q.shape[-1]
assert qkv_features % num_heads == 0, (
'Memory dimension must be divisible by number of heads.')
head_dim = qkv_features // num_heads
dense = nn.DenseGeneral.partial(axis=-1,
features=(num_heads, head_dim),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision)
# project inputs_q to multi-headed q/k/v
# dimensions are then [bs, dims..., n_heads, n_features_per_head]
qlength = inputs_q.shape[-2]
kvlength = inputs_kv.shape[-2]
if ignore_dot_product:
value = dense(inputs_kv, dtype=dtype, name='value')
key = value
query = inputs_q
else:
query, key, value = (dense(inputs_q, dtype=dtype, name='query'),
dense(inputs_kv, dtype=dtype, name='key'),
dense(inputs_kv, dtype=dtype, name='value'))
syn_weights_list = []
logging.info(synthesizer_mode)
if 'random' in synthesizer_mode:
if 'factorized_random' in synthesizer_mode:
logging.info('Using factorized random')
rand_syn_weights1 = self.param('random1',
(num_heads, max_length, k),
kernel_init)
rand_syn_weights2 = self.param('random2',
(num_heads, k, max_length),
kernel_init)
rand_syn_weights1 = rand_syn_weights1[:, :qlength, :]
rand_syn_weights2 = rand_syn_weights2[:, :, :kvlength]
rand_syn_weights = jnp.einsum('hlk,hkn->hln',
rand_syn_weights1,
rand_syn_weights2)
rand_syn_weights = jax.lax.broadcast(rand_syn_weights,
(inputs_q.shape[0], ))
syn_weights_list.append(rand_syn_weights)
else:
rand_syn_weights = self.param(
'random', (num_heads, max_length, max_length), kernel_init)
rand_syn_weights = rand_syn_weights[:, :qlength, :kvlength]
rand_syn_weights = jax.lax.broadcast(rand_syn_weights,
(inputs_q.shape[0], ))
syn_weights_list.append(rand_syn_weights)
if 'dense' in synthesizer_mode:
dense_syn = nn.DenseGeneral.partial(axis=-1,
features=(num_heads, head_dim),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision,
name='dense_syn',
dtype=dtype)
# TODO(yitay): Change this to nn.Dense and make sure it works
dense_syn_length = nn.linear.DenseGeneral.partial(
axis=-1,
features=(max_length),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision,
name='dense_syn2',
dtype=dtype)
proj = dense_syn(inputs_q, dtype=dtype, name='dense_syn')
proj = jax.nn.relu(proj)
proj = dense_syn_length(proj, dtype=dtype, name='dense_syn_len')
# TODO(yitay) check if this reshape is needed
dense_syn_weights = proj.reshape(
(inputs_q.shape[0], num_heads, qlength, max_length))
dense_syn_weights = dense_syn_weights[:, :, :, :qlength]
syn_weights_list.append(dense_syn_weights)
if cache:
assert isinstance(cache,
Cache), 'cache must be an instance of Cache'
if self.is_initializing():
cache.store(
onp.array((key.ndim, ) + key.shape[-2:], dtype=onp.int32))
else:
cache_entry = cache.retrieve(None)
expected_shape = list(cache_entry.key.shape[:-2])
for attn_dim in attention_axis:
expected_shape[attn_dim] = 1
expected_shape = tuple(expected_shape) + inputs_q.shape[-1:]
if expected_shape != inputs_q.shape:
raise ValueError('Invalid shape provided, '
'expected shape %s instead got %s.' %
(expected_shape, inputs_q.shape))
if not isinstance(cache_entry, _CacheEntry):
raise ValueError('Cache is not initialized.')
cshape = cache_entry.key.shape
indices = [0] * len(cshape)
i = cache_entry.i
attn_size = onp.prod(onp.take(cshape, attention_axis))
for attn_dim in attention_axis:
attn_size //= cshape[attn_dim]
indices[attn_dim] = i // attn_size
i = i % attn_size
key = lax.dynamic_update_slice(cache_entry.key, key, indices)
value = lax.dynamic_update_slice(cache_entry.value, value,
indices)
one = jnp.array(1, jnp.uint32)
cache_entry = cache_entry.replace(i=cache_entry.i + one,
key=key,
value=value)
cache.store(cache_entry)
key_padding_mask = jnp.broadcast_to(
(jnp.arange(cshape[1]) < cache_entry.i), cshape[:2])
key_padding_mask = key_padding_mask.astype(jnp.float32)[...,
None]
# create attention masks
mask_components = []
if causal_mask:
if cache and not self.is_initializing():
bias_pre_shape = (1, ) * (key.ndim - 1)
attn_shape = tuple(onp.take(key.shape, attention_axis))
attn_size = onp.prod(attn_shape)
ii = jnp.arange(attn_size, dtype=jnp.uint32)
mask = ii < cache_entry.i
mask_components.append(
mask.reshape(bias_pre_shape + attn_shape))
else:
mask_components.append(_make_causal_mask(key, attention_axis))
if not ignore_dot_product:
if padding_mask is not None:
if key_padding_mask is None:
key_padding_mask = padding_mask
padding_mask = make_padding_mask(
padding_mask_query=padding_mask,
padding_mask_key=key_padding_mask,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis)
mask_components.append(padding_mask)
if segmentation is not None:
if key_segmentation is None:
key_segmentation = segmentation
segmentation_mask = make_padding_mask(
padding_mask_query=segmentation,
padding_mask_key=key_segmentation,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis,
segmentation_mask=True)
mask_components.append(segmentation_mask)
if mask_components:
attention_mask = mask_components[0]
for component in mask_components[1:]:
attention_mask = jnp.logical_and(attention_mask, component)
# attention mask in the form of attention bias
attention_bias = lax.select(
attention_mask > 0,
jnp.full(attention_mask.shape, 0.).astype(dtype),
jnp.full(attention_mask.shape, -1e10).astype(dtype))
else:
attention_bias = None
# apply attention
x = synthetic_attention(query,
key,
value,
syn_weights_list,
dtype=dtype,
axis=attention_axis,
bias=attention_bias,
precision=precision,
dropout_rng=dropout_rng,
dropout_rate=dropout_rate,
broadcast_dropout=broadcast_dropout,
deterministic=deterministic,
ignore_dot_product=ignore_dot_product)
# back to the original inputs dimensions
out = nn.DenseGeneral(x,
features=features,
axis=(-2, -1),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
dtype=dtype,
precision=precision,
name='out')
return out
def apply(self,
inputs_q,
inputs_kv,
num_heads,
block_size=64,
num_rand_blocks=3,
dtype=jnp.float32,
qkv_features=None,
out_features=None,
attention_axis=None,
causal_mask=False,
padding_mask=None,
key_padding_mask=None,
segmentation=None,
key_segmentation=None,
cache=None,
broadcast_dropout=True,
dropout_rng=None,
dropout_rate=0.,
deterministic=False,
precision=None,
kernel_init=nn.linear.default_kernel_init,
bias_init=nn.initializers.zeros,
bias=True,
connectivity_seed=None):
"""Applies multi-head dot product attention on the input data.
Projects the inputs into multi-headed query, key, and value vectors,
applies dot-product attention and project the results to an output vector.
This can be used for encoder-decoder attention by specifying both `inputs_q`
and `inputs_kv` orfor self-attention by only specifying `inputs_q` and
setting `inputs_kv` to None.
Args:
inputs_q: input queries of shape `[bs, length, features]`.
inputs_kv: key/values of shape `[bs, length, features]` or
None for self-attention, inn which case key/values will be derived from
inputs_q.
num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])
should be divisible by the number of heads.
block_size: Size for local attention around diagonal of attention.
num_rand_blocks: int. Number of random chunks per row.
dtype: the dtype of the computation (default: float32)
qkv_features: dimension of the key, query, and value.
out_features: dimension of the last projection
attention_axis: axes over which the attention is applied ( 'None' means
attention over all axes, but batch, heads, and features).
causal_mask: boolean specifying whether to apply a causal mask on the
attention weights. If True, the output at timestep `t` will not depend
on inputs at timesteps strictly greater than `t`.
padding_mask: boolean specifying query tokens that are pad token.
key_padding_mask: boolean specifying key-value tokens that are pad token.
segmentation: segment indices for packed inputs_q data.
key_segmentation: segment indices for packed inputs_kv data.
cache: an instance of `flax.nn.attention.Cache` used for efficient
autoregressive decoding.
broadcast_dropout: bool: use a broadcasted dropout along batch dims.
dropout_rng: JAX PRNGKey: to be used for dropout
dropout_rate: dropout rate
deterministic: bool, deterministic or not (to apply dropout)
precision: numerical precision of the computation see `jax.lax.Precision`
for details.
kernel_init: initializer for the kernel of the Dense layers.
bias_init: initializer for the bias of the Dense layers.
bias: bool: whether pointwise QKVO dense transforms use bias.
connectivity_seed: Seed for random block sparse attention.
Returns:
output of shape `[bs, length, features]`.
"""
orig_seqlen = inputs_q.shape[-2]
logging.info(inputs_q)
extra_len = block_size - (orig_seqlen % block_size)
pad_width = jnp.array([[0, 0], [0, extra_len], [0, 0]])
mask_pad = jnp.array([[0, 0], [0, extra_len], [0, 0]])
padding_mask = jnp.pad(padding_mask, mask_pad, constant_values=-1e9)
inputs_q = jnp.pad(inputs_q, pad_width)
if inputs_kv is not None:
inputs_kv = jnp.pad(inputs_kv, pad_width)
assert causal_mask or not cache, (
'Caching is only support for causal attention.')
if inputs_kv is None:
inputs_kv = inputs_q
if attention_axis is None:
attention_axis = tuple(range(1, inputs_q.ndim - 1))
features = out_features or inputs_q.shape[-1]
qkv_features = qkv_features or inputs_q.shape[-1]
assert qkv_features % num_heads == 0, (
'Memory dimension must be divisible by number of heads.')
head_dim = qkv_features // num_heads
dense = nn.DenseGeneral.partial(axis=-1,
features=(num_heads, head_dim),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
precision=precision)
# project inputs_q to multi-headed q/k/v
# dimensions are then [bs, dims..., n_heads, n_features_per_head]
query, key, value = (dense(inputs_q, dtype=dtype, name='query'),
dense(inputs_kv, dtype=dtype, name='key'),
dense(inputs_kv, dtype=dtype, name='value'))
if cache:
assert isinstance(
cache, attention.Cache), 'cache must be an instance of Cache'
if self.is_initializing():
cache.store(
onp.array((key.ndim, ) + key.shape[-2:], dtype=onp.int32))
else:
cache_entry = cache.retrieve(None)
expected_shape = list(cache_entry.key.shape[:-2])
for attn_dim in attention_axis:
expected_shape[attn_dim] = 1
expected_shape = tuple(expected_shape) + inputs_q.shape[-1:]
if expected_shape != inputs_q.shape:
raise ValueError('Invalid shape provided, '
'expected shape %s instead got %s.' %
(expected_shape, inputs_q.shape))
if not isinstance(cache_entry, attention._CacheEntry): # pylint: disable=protected-access
raise ValueError('Cache is not initialized.')
cshape = cache_entry.key.shape
indices = [0] * len(cshape)
i = cache_entry.i
attn_size = onp.prod(onp.take(cshape, attention_axis))
for attn_dim in attention_axis:
attn_size //= cshape[attn_dim]
indices[attn_dim] = i // attn_size
i = i % attn_size
key = lax.dynamic_update_slice(cache_entry.key, key, indices)
value = lax.dynamic_update_slice(cache_entry.value, value,
indices)
one = jnp.array(1, jnp.uint32)
cache_entry = cache_entry.replace(i=cache_entry.i + one,
key=key,
value=value)
cache.store(cache_entry)
# TODO(levskaya): verify this is still needed in translation decoding.
key_padding_mask = jnp.broadcast_to(
(jnp.arange(cshape[1]) < cache_entry.i), cshape[:2])
key_padding_mask = key_padding_mask.astype(jnp.float32)[...,
None]
if causal_mask: # Falls back to full attention with a causal mask.
# create attention masks
mask_components = []
if causal_mask:
if cache and not self.is_initializing():
bias_pre_shape = (1, ) * (key.ndim - 1)
attn_shape = tuple(onp.take(key.shape, attention_axis))
attn_size = onp.prod(attn_shape)
ii = jnp.arange(attn_size, dtype=jnp.uint32)
mask = ii < cache_entry.i
mask_components.append(
mask.reshape(bias_pre_shape + attn_shape))
else:
mask_components.append(
attention._make_causal_mask(key, attention_axis)) # pylint: disable=protected-access
if padding_mask is not None:
if key_padding_mask is None:
key_padding_mask = padding_mask
padding_mask = attention.make_padding_mask(
padding_mask_query=padding_mask,
padding_mask_key=key_padding_mask,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis)
mask_components.append(padding_mask)
if segmentation is not None:
if key_segmentation is None:
key_segmentation = segmentation
segmentation_mask = attention.make_padding_mask(
padding_mask_query=segmentation,
padding_mask_key=key_segmentation,
query_shape=query.shape,
key_shape=key.shape,
attention_axis=attention_axis,
segmentation_mask=True)
mask_components.append(segmentation_mask)
if mask_components:
attention_mask = mask_components[0]
for component in mask_components[1:]:
attention_mask = jnp.logical_and(attention_mask, component)
# attention mask in the form of attention bias
attention_bias = lax.select(
attention_mask > 0,
jnp.full(attention_mask.shape, 0.).astype(dtype),
jnp.full(attention_mask.shape, -1e10).astype(dtype))
else:
attention_bias = None
x = nn.dot_product_attention(query,
key,
value,
dtype=dtype,
axis=attention_axis,
bias=attention_bias,
precision=precision,
dropout_rng=dropout_rng,
dropout_rate=dropout_rate,
broadcast_dropout=broadcast_dropout,
deterministic=deterministic)
else:
if connectivity_seed is None:
path = self._get_construction_frame().path
connectivity_seed = hash(path) % 2**32
# apply attention
input_mask = None
if padding_mask is not None:
input_mask = padding_mask.astype(key.dtype)
x = sparse_dot_product_attention(
query,
key,
value,
connectivity_seed=connectivity_seed,
input_mask=input_mask,
block_size=block_size,
num_rand_blocks=num_rand_blocks)
# back to the original inputs dimensions
out = nn.DenseGeneral(x,
features=features,
axis=(-2, -1),
kernel_init=kernel_init,
bias_init=bias_init,
bias=bias,
dtype=dtype,
precision=precision,
name='out')
out = out[:, :orig_seqlen, :]
return out