def _noisy_targets(self, targets, losses=None):
"""Generate noisy targets for denoising models.
Args:
targets: a Tensor
losses: an optional list onto which to append traning losses
Returns:
a Tensor the same dtype and shape as Targets
"""
hparams = self._hparams
if hparams.mode == tf.estimator.ModeKeys.TRAIN:
nt_train = self._noisy_targets_from_spec(
targets, hparams.noising_spec_train, losses=losses)
if hparams.noising_use_eval_during_train > 0:
nt_eval = self._noisy_targets_from_spec(
targets, hparams.noising_spec_eval)
use_eval_noising = mtf.less(
mtf.random_uniform(targets.mesh,
targets.shape - self.length_dim),
hparams.noising_use_eval_during_train)
nt_train = mtf.where(use_eval_noising, nt_eval, nt_train)
return nt_train
else:
return self._noisy_targets_from_spec(targets,
hparams.noising_spec_eval)
def compute_mask(self, context, memory_position):
"""Compute attention mask.
Args:
context: a transformer.Context
memory_position: an int32 tensor containing memory_length dimension.
Returns:
a Tensor or None
"""
masks = []
min_relative_position = self.min_relative_position(context)
max_relative_position = self.max_relative_position(context)
if max_relative_position is not None or min_relative_position is not None:
relative_position = memory_position - context.position
if min_relative_position is not None:
illegal = mtf.less(relative_position, min_relative_position)
masks.append(mtf.cast(illegal, context.activation_dtype) * -1e9)
if max_relative_position is not None:
illegal = mtf.greater(relative_position, max_relative_position)
masks.append(mtf.cast(illegal, context.activation_dtype) * -1e9)
if (context.sequence_id is not None and
isinstance(context.sequence_id, mtf.Tensor) and
context.length_dim in context.sequence_id.shape):
masks.append(mtf.cast(
mtf.not_equal(
context.sequence_id,
self.rename_length_to_memory_length(
context.sequence_id, context)),
context.activation_dtype) * -1e9)
return mtf.add_n(masks) if masks else None
def get_attn_mask(self, mesh, nd, ns):
if not exists(self.attn_mask):
i = mtf.range(mesh, nd, tf.int32) + ns.size - nd.size
j = mtf.range(mesh, ns, tf.int32)
i, j = map(lambda t: mtf.broadcast(t, [nd, ns]), (i, j))
self.attn_mask = mtf.cast(mtf.less(
i, j), self.variable_dtype.activation_dtype) * -1e10
return self.attn_mask
def _relative_position_bucket(relative_position,
bidirectional=True,
num_buckets=32,
max_distance=128):
"""Translate relative position to a bucket number for relative attention.
The relative position is defined as memory_position - query_position, i.e.
the distance in tokens from the attending position to the attended-to
position. If bidirectional=False, then positive relative positions are
invalid.
We use smaller buckets for small absolute relative_position and larger buckets
for larger absolute relative_positions. All relative positions >=max_distance
map to the same bucket. All relative positions <=-max_distance map to the
same bucket. This should allow for more graceful generalization to longer
sequences than the model has been trained on.
Args:
relative_position: an int32 Tensor
bidirectional: a boolean - whether the attention is bidirectional
num_buckets: an integer
max_distance: an integer
Returns:
a Tensor with the same shape as relative_position, containing int32
values in the range [0, num_buckets)
"""
ret = 0
n = -relative_position
if bidirectional:
num_buckets //= 2
ret += mtf.to_int32(mtf.less(n, 0)) * num_buckets
n = mtf.abs(n)
else:
n = mtf.maximum(n, 0)
# now n is in the range [0, inf)
max_exact = num_buckets // 2
is_small = mtf.less(n, max_exact)
val_if_large = max_exact + mtf.to_int32(
mtf.log(mtf.to_float(n) / max_exact)
/ math.log(max_distance / max_exact) * (num_buckets - max_exact))
val_if_large = mtf.minimum(val_if_large, num_buckets - 1)
ret += mtf.where(is_small, n, val_if_large)
return ret
def get_log_softmax_prefix(self, log_softmax, end_index):
"""Returns first end_index entries in log_softmax along the vocab dim."""
prefix_dim = mtf.Dimension(self._vocab_dim.name, end_index)
indices = mtf.mtf_range(
log_softmax.mesh, dim=self._vocab_dim, dtype=tf.int32)
prefix_indices = mtf.where(mtf.less(indices, end_index), indices, -1)
projection = mtf.one_hot(
prefix_indices, prefix_dim, dtype=log_softmax.dtype)
return mtf.einsum([log_softmax, projection], reduced_dims=[self._vocab_dim])
def biasmask_attn_weights(mesh, nd, ns, variable_dtype):
# The old mask_attn_weights applied directly to the QK;
# this returns a bias that the attention code from mtf adds to the attention matrix.
# w has shape [batch, heads, dst_sequence, src_sequence], where information flows from src to dst.
# n_src and n_dest are both the same, i.e equal to sequence length
# We rename ns because we want bias to have shape [batch, heads, memory_length, sequence] to match up with QK^T
# Information flows from k and v (memory_length) to q (sequence)
i = mtf.range(mesh, nd, tf.int32) + ns.size - nd.size
j = mtf.range(mesh, ns, tf.int32)
i, j = map(lambda t: mtf.broadcast(t, [nd, ns]), (i, j))
dtype = variable_dtype.activation_dtype
return mtf.cast(mtf.less(i, j), dtype) * -1e10
def testWhileLoopOperation(self):
# This test case implements the following:
# for i in range(10):
# x = x * 2
i = mtf.constant(self.mesh, 0, mtf.Shape([]))
cond_fn = lambda i, x: mtf.less(i, 10)
body_fn = lambda i, x: [mtf.add(i, 1), mtf.multiply(x, 2)]
while_loop_operation = mtf.WhileLoopOperation(cond_fn, body_fn, [i, self.x])
self.assertEqual(while_loop_operation.splittable_dims,
frozenset(["a", "b"]))
self.assertEqual(while_loop_operation.unsplittable_dims, frozenset())
def call(self, context, x, losses=None):
"""Call the layer."""
wq, wk, wv, wo = mtf.layers.multihead_attention_params(
context.mesh, self.heads_dim, context.model_dim, self.kv_dim,
context.variable_dtype)
memory_length = mtf.Dimension("memory_length", context.length_dim.size)
q = mtf.einsum([x, wq], reduced_dims=[context.model_dim])
if context.mode == "incremental":
m = x
else:
m = mtf.rename_dimension(x, context.length_dim.name,
"memory_length")
k = mtf.einsum([m, wk], reduced_dims=[context.model_dim])
v = mtf.einsum([m, wv], reduced_dims=[context.model_dim])
if context.mode == "incremental":
old_k, old_v = context.get_states(2)
one_hot = mtf.one_hot(context.position,
memory_length,
dtype=context.activation_dtype)
inv_one_hot = 1.0 - one_hot
k = old_k * inv_one_hot + k * one_hot
v = old_v * inv_one_hot + v * one_hot
if context.mode == "incremental" or context.mode == "first_part":
context.record_new_states([k, v])
masks = []
if context.autoregressive:
masks.append(
mtf.cast(
mtf.less(
context.position,
mtf.range(context.mesh, memory_length,
dtype=tf.int32)), context.activation_dtype) *
-1e9)
if (context.sequence_id is not None
and isinstance(context.sequence_id, mtf.Tensor)
and context.length_dim in context.sequence_id.shape):
masks.append(
mtf.cast(
mtf.not_equal(
context.sequence_id,
mtf.layers.rename_length_to_memory_length(
context.sequence_id)), context.activation_dtype) *
-1e9)
mask = mtf.add_n(masks) if masks else None
o = mtf.layers.dot_product_attention_v2(
q, k, v, memory_length, self.kv_dim, self.kv_dim, mask,
self.dropout_rate if context.train else 0.0, [context.length_dim])
return mtf.einsum([o, wo],
x.shape,
reduced_dims=[self.heads_dim, self.kv_dim])
def ids_to_embedding(self, ids, context):
"""Ids to embeddings with ids not in cluster mapped to the zero vector."""
ids -= self._start_token_id
# The mtf.gather in the embedding's ids_to_embedding implementation will
# cause the one hot representations of tokens greater than cluster vocab
# dimension size to be the zero vector. Thus the embeddings for those tokens
# will be the zero vector.
ids = mtf.where(mtf.greater_equal(ids, 0), ids, self._vocab_dim.size)
# Handle the case of the head cluster where we will have entries at the end
# corresponding to the tail clusters.
ids = mtf.where(
mtf.less(ids, self._end_token_id - self._start_token_id),
ids,
self._vocab_dim.size,
)
return self._embedding.ids_to_embedding(ids, context)
def call(self, context, x, losses=None):
"""Call the layer."""
params = mtf.layers.multihead_attention_params(context.mesh,
self.heads_dim,
context.model_dim,
self.kv_dim,
context.variable_dtype)
if context.mode == "incremental":
prev_k, prev_v = context.get_states(2)
y, new_k, new_v = mtf.layers.masked_local_attention_1d_incremental(
x, prev_k, prev_v, context.position, params=params)
context.record_new_states([new_k, new_v])
return y
else:
kv = []
y = mtf.layers.masked_local_attention_1d(x,
self.kv_dim,
self.heads_dim,
self.window_size,
params=params,
return_kv=kv)
if context.mode == "first_part":
k = kv[0]
v = kv[1]
window_dim = mtf.Dimension("window", self.window_size)
mesh = k.mesh
window_pos = mtf.range(mesh, window_dim, tf.int32)
pos = mtf.range(mesh, context.length_dim, tf.int32)
select_recent = mtf.cast(
mtf.equal(window_pos, mtf.mod(pos, self.window_size)),
k.dtype)
select_recent *= mtf.cast(
mtf.less(pos, context.initial_position), k.dtype)
select_recent *= mtf.cast(
mtf.greater_equal(
pos, context.initial_position - self.window_size),
k.dtype)
state_shape = k.shape.dims[:-2] + [window_dim, self.kv_dim]
k_state = mtf.einsum([k, select_recent],
output_shape=state_shape,
reduced_dims=[context.length_dim])
v_state = mtf.einsum([v, select_recent],
output_shape=state_shape,
reduced_dims=[context.length_dim])
context.new_states.extend([k_state, v_state])
return y
def _noisy_targets_from_spec(self, targets, noising_spec, losses=None):
if noising_spec["type"] == "mask":
# Replace a randomly-chosen noising_spec["prob"] of input tokens with 0.
return targets * mtf.cast(
mtf.greater(mtf.random_uniform(targets.mesh, targets.shape),
noising_spec["prob"]), targets.dtype)
elif noising_spec["type"] == "random_zipfian":
# Replace a randomly-chosen noising_spec["prob"] of input tokens.
# Rather than drawing the replacement tokens uniformly, we sample from
# a distribution favoring lower token-ids, assuming that the ids have
# been assigned in frequency order. The probability of choosing an
# id is proportional to 1/(id+10)
logits = mtf.log(1.0 / (mtf.range(
targets.mesh, self.targets_vocab_dim, dtype=tf.float32) +
10.0))
logits = mtf.broadcast(logits,
new_shape=targets.shape + logits.shape)
r = mtf.sample_with_temperature(logits, self.targets_vocab_dim)
use_noise = mtf.less(
mtf.random_uniform(targets.mesh, targets.shape),
noising_spec["prob"])
return mtf.where(use_noise, r, targets)
elif noising_spec["type"] == "transformer":
# Train a small transformer to fill in masked out values, then
# sample from it.
hparams = self._hparams
if hparams.mode != tf.estimator.ModeKeys.TRAIN:
raise NotImplementedError("Not implemented")
noiser_hparams = copy.copy(self._hparams)
noiser_hparams.del_hparam("mode")
noiser_hparams.override_from_dict(noising_spec["overrides"])
with tf.variable_scope("noiser"):
noiser = MtfTransformer(noiser_hparams,
mode=hparams.mode,
problem_hparams=self._problem_hparams)
logits, loss = noiser._mtf_model_fn( # pylint: disable=protected-access
self._original_features, targets.mesh)
samples = mtf.sample_with_temperature(logits,
self.targets_vocab_dim)
losses.append(loss)
return samples
else:
raise ValueError("unknown noising spec %s" % noising_spec)
def attention(x, dim_head, dim_features_head, scope='attn', causal=False):
with tf.variable_scope(scope):
mesh, batch, seq, dim = x.mesh, *x.shape
dim_heads = mtf.Dimension('dim_heads',
dim_head.size * dim_features_head.size)
dim_intermediate = mtf.Dimension('qkv_dimension', dim_heads.size * 3)
qkv = linear(x, dim_intermediate, bias=False, scope='to_qkv')
q, k, v = mtf.split(qkv, dim_intermediate, 3)
q, k, v = map(
lambda t: mtf.reshape(t, [batch, seq, dim_head, dim_features_head]
), (q, k, v))
q, k, v = map(
lambda t: mtf.transpose(
t, [batch, dim_head, seq, dim_features_head]), (q, k, v))
k, v = map(
lambda t: mtf.rename_dimension(t, seq.name, 'memory_length'),
(k, v))
mem_len_dim = v.shape[-2]
dots = mtf.layers.us_einsum([q, k],
[batch, dim_head, seq, mem_len_dim])
if causal:
i = mtf.range(mesh, seq, tf.int32)
j = mtf.range(mesh, mem_len_dim, tf.int32)
i, j = map(lambda t: mtf.broadcast(t, [seq, mem_len_dim]), (i, j))
mask = mtf.less(i + mem_len_dim.size - seq.size, j)
mask = mtf.cast(mask, tf.float32) * -1e10
dots += mask
attn = mtf.softmax(dots, mem_len_dim)
out = mtf.einsum([attn, v], [batch, dim_head, seq, dim_features_head])
out = mtf.transpose(out, [batch, seq, dim_head, dim_features_head])
out = mtf.reshape(out, [batch, seq, dim_heads])
combined_out = linear(out, dim, scope='combine_output')
return combined_out
def _noisy_targets_from_spec(self, targets, noising_spec, losses=None):
if noising_spec["type"] == "mask":
# Replace a randomly-chosen noising_spec["prob"] of input tokens with 0.
return targets * mtf.cast(
mtf.greater(mtf.random_uniform(targets.mesh, targets.shape),
noising_spec["prob"]), targets.dtype)
elif noising_spec["type"] == "random_zipfian":
# Replace a randomly-chosen noising_spec["prob"] of input tokens.
# Rather than drawing the replacement tokens uniformly, we sample from
# a distribution favoring lower token-ids, assuming that the ids have
# been assigned in frequency order. The probability of choosing an
# id is proportional to 1/(id+10)
logits = mtf.log(1.0 / (mtf.range(
targets.mesh, self.targets_vocab_dim, dtype=tf.float32) + 10.0))
logits = mtf.broadcast(logits, new_shape=targets.shape + logits.shape)
r = mtf.sample_with_temperature(logits, self.targets_vocab_dim)
use_noise = mtf.less(
mtf.random_uniform(targets.mesh, targets.shape), noising_spec["prob"])
return mtf.where(use_noise, r, targets)
elif noising_spec["type"] == "transformer":
# Train a small transformer to fill in masked out values, then
# sample from it.
hparams = self._hparams
if hparams.mode != tf.estimator.ModeKeys.TRAIN:
raise NotImplementedError("Not implemented")
noiser_hparams = copy.copy(self._hparams)
noiser_hparams.del_hparam("mode")
noiser_hparams.override_from_dict(noising_spec["overrides"])
with tf.variable_scope("noiser"):
noiser = MtfTransformer(
noiser_hparams,
mode=hparams.mode,
problem_hparams=self._problem_hparams)
logits, loss = noiser._mtf_model_fn( # pylint: disable=protected-access
self._original_features, targets.mesh)
samples = mtf.sample_with_temperature(logits, self.targets_vocab_dim)
losses.append(loss)
return samples
else:
raise ValueError("unknown noising spec %s" % noising_spec)
def compute_mask(self, context, memory_position):
"""Compute attention mask.
Args:
context: a transformer.Context
memory_position: an int32 tensor containing memory_length dimension.
Returns:
a Tensor or None
"""
masks = []
min_relative_position = self.min_relative_position(context)
max_relative_position = self.max_relative_position(context)
if max_relative_position is not None or min_relative_position is not None:
relative_position = memory_position - context.position
if min_relative_position is not None:
illegal = mtf.less(relative_position, min_relative_position)
masks.append(
mtf.cast(illegal, context.activation_dtype) * -1e9)
if max_relative_position is not None:
illegal = mtf.greater(relative_position, max_relative_position)
masks.append(
mtf.cast(illegal, context.activation_dtype) * -1e9)
sequence_id = None
# Subsequence id should only be set if we are in the decoder and have
# multiple targets per input. This will allow each sub-target to only attend
# to itself.
if isinstance(context.subsequence_id, mtf.Tensor):
sequence_id = context.subsequence_id
elif isinstance(context.sequence_id, mtf.Tensor):
sequence_id = context.sequence_id
if (sequence_id is not None
and context.length_dim in sequence_id.shape):
masks.append(
mtf.cast(
mtf.not_equal(
sequence_id,
self.rename_length_to_memory_length(
sequence_id, context)), context.activation_dtype) *
-1e9)
return mtf.add_n(masks) if masks else None
def _noisy_targets(self, targets, losses=None):
"""Generate noisy targets for denoising models.
Args:
targets: a Tensor
losses: an optional list onto which to append traning losses
Returns:
a Tensor the same dtype and shape as Targets
"""
hparams = self._hparams
if hparams.mode == tf.estimator.ModeKeys.TRAIN:
nt_train = self._noisy_targets_from_spec(
targets, hparams.noising_spec_train, losses=losses)
if hparams.noising_use_eval_during_train > 0:
nt_eval = self._noisy_targets_from_spec(
targets, hparams.noising_spec_eval)
use_eval_noising = mtf.less(
mtf.random_uniform(targets.mesh, targets.shape - self.length_dim),
hparams.noising_use_eval_during_train)
nt_train = mtf.where(use_eval_noising, nt_eval, nt_train)
return nt_train
else:
return self._noisy_targets_from_spec(targets, hparams.noising_spec_eval)
def _top_2_gating(inputs,
outer_expert_dims,
experts_dim,
expert_capacity_dim,
hparams,
train,
variable_dtype,
importance=None,
name="top_2_gating"):
"""Compute gating for mixture-of-experts in TensorFlow.
Note: until the algorithm and inferface solidify, we pass in a hyperparameters
dictionary in order not to complicate the interface in mtf_transformer.py .
Once this code moves out of "research", we should pass the hyperparameters
separately.
Hyperparameters used:
hparams.moe_use_second_place_loss: a boolean
hparams.moe_second_policy_train: a string
hparams.moe_second_policy_eval: a string
hparams.moe_second_threshold: a float
The returned forward assignment is a tensor used to map (via einsum) from the
inputs to the expert_inputs. Likewise, the returned combine_tensor is
used to map (via einsum) from the expert outputs to the outputs. Both the
forward and backward assignments are mostly zeros. The shapes of the tensors
are as follows.
inputs: [<batch_dims>, group_size_dim, input_dim]
importance: [<batch_dims>, group_size_dim]
dispatch_tensor:
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
expert_inputs:
[<batch_dims>, experts_dim, expert_capacity_dim, input_dim]
expert_outputs: [<batch_dims>, experts_dim, expert_capacity_dim, output_dim]
combine_tensor:
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
outputs: [<batch_dims>, group_size_dim, output_dim]
"importance" is an optional tensor with one floating-point value for each
input vector. If the importance of an input is 1.0, then we send it to
up to 2 experts. If 0.0 < importance < 1.0, then we send it to at most
one expert. If importance == 0.0, then we send it to no experts.
We use "importance" at the second-level gating function of a hierarchical
mixture of experts. Inputs to the first-choice expert-group get importance
1.0. Inputs to the second-choice expert group get importance 0.5.
Inputs that represent padding get importance 0.0.
Args:
inputs: a mtf.Tensor with shape [<batch_dims>, group_size_dim, input_dim]
outer_expert_dims: an optional list of dimensions. This is for the case
where we are at an inner level of a hierarchical MoE.
experts_dim: a Dimension (the number of experts)
expert_capacity_dim: a Dimension (number of examples per group per expert)
hparams: model hyperparameters.
train: a boolean
variable_dtype: a mtf.VariableDType
importance: an optional tensor with shape [<batch_dims>, group_size_dim]
name: an optional string
Returns:
dispatch_tensor: a Tensor with shape
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
combine_tensor: a Tensor with shape
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
loss: a mtf scalar
Raises:
ValueError: on illegal hyperparameters
"""
group_size_dim, unused_input_dim = inputs.shape.dims[-2:]
raw_gates = mtf.layers.dense(inputs,
experts_dim,
use_bias=False,
expert_dims=outer_expert_dims,
variable_dtype=variable_dtype,
name=name)
raw_gates = mtf.softmax(raw_gates, experts_dim)
# The internals of this function run in float32.
# bfloat16 seems to reduce quality.
raw_gates = mtf.to_float(raw_gates)
expert_capacity_f = float(expert_capacity_dim.size)
# FIND TOP 2 EXPERTS PER POSITON
# Find the top expert for each position. shape=[batch, group]
index_1, gate_1 = mtf.top_1(raw_gates, experts_dim)
# [batch, group, experts]
mask_1 = mtf.one_hot(index_1, experts_dim, dtype=raw_gates.dtype)
density_1_proxy = raw_gates
if importance is not None:
mask_1 *= mtf.to_float(mtf.equal(importance, 1.0))
gate_1 *= mtf.to_float(mtf.equal(importance, 1.0))
density_1_proxy *= mtf.to_float(mtf.equal(importance, 1.0))
gates_without_top_1 = raw_gates * (1.0 - mask_1)
# [batch, group]
index_2, gate_2 = mtf.top_1(gates_without_top_1, experts_dim)
# [batch, group, experts]
mask_2 = mtf.one_hot(index_2, experts_dim, dtype=raw_gates.dtype)
if importance is not None:
mask_2 *= mtf.to_float(mtf.greater(importance, 0.0))
denom = gate_1 + gate_2 + 1e-9
gate_1 /= denom
gate_2 /= denom
# BALANCING LOSSES
# shape = [batch, experts]
# We want to equalize the fraction of the batch assigned to each expert
density_1 = mtf.reduce_mean(mask_1, reduced_dim=group_size_dim)
# Something continuous that is correlated with what we want to equalize.
density_1_proxy = mtf.reduce_mean(density_1_proxy,
reduced_dim=group_size_dim)
loss = (mtf.reduce_mean(density_1_proxy * density_1) *
float(experts_dim.size * experts_dim.size))
if hparams.moe_use_second_place_loss:
# Also add a loss to encourage all experts to be used equally also as the
# second-place expert. Experimentally, this seems to be a wash.
# We want to equalize the fraction of the batch assigned to each expert:
density_2 = mtf.reduce_mean(mask_2, reduced_dim=group_size_dim)
# As a proxy for density_2, we renormalize the raw gates after the top one
# has been removed.
normalized = gates_without_top_1 / (mtf.reduce_sum(
gates_without_top_1, reduced_dim=experts_dim) + 1e-9)
density_2_proxy = mtf.reduce_mean(normalized,
reduced_dim=group_size_dim)
loss_2 = (mtf.reduce_mean(density_2_proxy * density_2) *
float(experts_dim.size * experts_dim.size))
loss += loss_2 * 0.5
# Depending on the policy in the hparams, we may drop out some of the
# second-place experts.
if train:
policy = hparams.moe_second_policy_train
threshold = hparams.moe_second_threshold_train
else:
policy = hparams.moe_second_policy_eval
threshold = hparams.moe_second_threshold_eval
if policy == "all":
# Use second-place experts for all examples.
pass
elif policy == "none":
# Never use second-place experts for all examples.
mask_2 = mtf.zeros_like(mask_2)
elif policy == "threshold":
# Use second-place experts if gate_2 > threshold.
mask_2 *= mtf.to_float(mtf.greater(gate_2, threshold))
elif policy == "random":
# Use second-place experts with probablity min(1.0, gate_2 / threshold).
mask_2 *= mtf.to_float(
mtf.less(mtf.random_uniform(gate_2.mesh, gate_2.shape),
gate_2 / max(threshold, 1e-9)))
else:
raise ValueError("Unknown policy %s" % policy)
# COMPUTE ASSIGNMENT TO EXPERTS
# [batch, group, experts]
# This is the position within the expert's mini-batch for this sequence
position_in_expert_1 = mtf.cumsum(mask_1, group_size_dim,
exclusive=True) * mask_1
# Remove the elements that don't fit. [batch, group, experts]
mask_1 *= mtf.to_float(mtf.less(position_in_expert_1, expert_capacity_f))
# [batch, experts]
# How many examples in this sequence go to this expert
mask_1_count = mtf.reduce_sum(mask_1, reduced_dim=group_size_dim)
# [batch, group] - mostly ones, but zeros where something didn't fit
mask_1_flat = mtf.reduce_sum(mask_1, reduced_dim=experts_dim)
# [batch, group]
position_in_expert_1 = mtf.reduce_sum(position_in_expert_1,
reduced_dim=experts_dim)
# Weight assigned to first expert. [batch, group]
gate_1 *= mask_1_flat
# [batch, group, experts]
position_in_expert_2 = (
mtf.cumsum(mask_2, group_size_dim, exclusive=True) + mask_1_count)
position_in_expert_2 *= mask_2
mask_2 *= mtf.to_float(mtf.less(position_in_expert_2, expert_capacity_f))
# mask_2_count = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
mask_2_flat = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
gate_2 *= mask_2_flat
position_in_expert_2 = mtf.reduce_sum(position_in_expert_2,
reduced_dim=experts_dim)
# [batch, group, experts, expert_capacity]
combine_tensor = (
gate_1 * mask_1_flat * mtf.one_hot(index_1, experts_dim) *
mtf.one_hot(mtf.to_int32(position_in_expert_1), expert_capacity_dim) +
gate_2 * mask_2_flat * mtf.one_hot(index_2, experts_dim) *
mtf.one_hot(mtf.to_int32(position_in_expert_2), expert_capacity_dim))
combine_tensor = mtf.cast(combine_tensor, inputs.dtype)
loss = mtf.cast(loss, inputs.dtype)
dispatch_tensor = mtf.cast(mtf.cast(combine_tensor, tf.bool),
combine_tensor.dtype)
return dispatch_tensor, combine_tensor, loss
def get_cluster_mask(self, targets):
"""Computes mask over the targets masking out tokens not in the cluster."""
return mtf.logical_and(
mtf.greater_equal(targets, self._start_token_id),
mtf.less(targets, self._end_token_id))
def _rand_1_gating(
inputs, outer_expert_dims, experts_dim, expert_capacity_dim,
hparams, train, variable_dtype, importance=None, name="rand_1_gating",
num_microbatches=None):
"""Compute a random top-1 gating."""
# SELECT EXPERT
if train:
policy = hparams.moe_rand_1_policy_train
else:
policy = hparams.moe_rand_1_policy_eval
# The internals of this function run in float32.
# bfloat16 seems to reduce quality.
gate_inputs = mtf.to_float(inputs)
# Input perturbations
if train and policy == "input_dropout":
gate_inputs = mtf.dropout(gate_inputs, 1.0 - hparams.moe_rand_1_dropout)
elif train and policy == "input_jitter":
gate_inputs = mtf.layers.multiplicative_jitter(gate_inputs,
hparams.moe_rand_1_jitter)
gate_logits = mtf.layers.dense(
gate_inputs,
experts_dim,
use_bias=False,
expert_dims=outer_expert_dims,
variable_dtype=variable_dtype,
name=name)
raw_gates = mtf.softmax(gate_logits, reduced_dim=experts_dim)
if policy == "argmax" or policy == "input_dropout" or policy == "input_jitter":
expert_gate, expert_index = mtf.top_1(raw_gates, reduced_dim=experts_dim)
elif policy == "sample":
expert_index = mtf.sample_with_temperature(
gate_logits, experts_dim, temperature=hparams.moe_rand_1_temperature)
expert_gate = mtf.gather(raw_gates, expert_index, dim=experts_dim)
else:
raise ValueError("Unknown rand_1 policy %s" % policy)
expert_mask = mtf.one_hot(expert_index, experts_dim, dtype=raw_gates.dtype)
# LOAD BALANCING LOSS
# TODO(liamfedus): Check entropy loss.
group_size_dim = inputs.shape[-2]
density_1 = mtf.reduce_mean(expert_mask, reduced_dim=group_size_dim)
density_1_proxy = mtf.reduce_mean(raw_gates, reduced_dim=group_size_dim)
if importance is not None:
expert_mask *= mtf.cast(mtf.equal(importance, 1.0), dtype=raw_gates.dtype)
expert_gate *= mtf.cast(mtf.equal(importance, 1.0), dtype=raw_gates.dtype)
density_1_proxy *= mtf.cast(
mtf.equal(importance, 1.0), dtype=raw_gates.dtype)
loss = (
mtf.reduce_mean(density_1_proxy * density_1) *
float(experts_dim.size * experts_dim.size))
if num_microbatches and num_microbatches > 1:
tf.logging.info("Dividing load-balance loss by num_microbatches={}".format(
num_microbatches))
loss /= num_microbatches
# Logging
if train:
entropy = mtf.reduce_sum(-raw_gates * mtf.log(raw_gates + 1e-9),
reduced_dim=experts_dim)
batch_entropy = mtf.reduce_mean(entropy)
mtf.scalar_summary(name + "/entropy", batch_entropy)
mask_count_experts = mtf.reduce_sum(expert_mask, output_shape=[experts_dim])
total_routed = mtf.reduce_sum(mask_count_experts)
expert_fraction = mtf.to_float(mask_count_experts / total_routed)
split_fractions = mtf.split(
expert_fraction,
split_dim=experts_dim,
num_or_size_splits=experts_dim.size)
for fraction in split_fractions:
mtf.scalar_summary("experts/" + fraction.name.replace(":", "/"),
mtf.reduce_mean(fraction))
mtf.scalar_summary("aux_loss", mtf.reduce_mean(loss))
# COMPUTE ASSIGNMENT TO EXPERT
# Experts have a limited capacity, ensure we do not exceed it. Construct
# the batch indices, to each expert, with position_in_expert
position_in_expert = mtf.cumsum(
expert_mask, group_size_dim, exclusive=True) * expert_mask
position_in_expert = mtf.cast(position_in_expert, dtype=raw_gates.dtype)
# Keep only tokens that fit within expert_capacity.
expert_capacity_float = float(expert_capacity_dim.size)
expert_mask *= mtf.cast(
mtf.less(position_in_expert, expert_capacity_float),
dtype=raw_gates.dtype)
expert_mask_flat = mtf.reduce_sum(expert_mask, reduced_dim=experts_dim)
# Mask out the experts that have overflowed expert capacity. Sparsify the
# expert_gate.
expert_gate *= expert_mask_flat
combine_tensor = (
expert_gate * expert_mask_flat *
mtf.one_hot(expert_index, experts_dim, dtype=raw_gates.dtype) *
mtf.one_hot(
mtf.to_int32(position_in_expert),
expert_capacity_dim,
dtype=raw_gates.dtype))
# Match the inputs dtype.
combine_tensor = mtf.cast(combine_tensor, inputs.dtype)
loss = mtf.cast(loss, inputs.dtype)
dispatch_tensor = mtf.cast(
mtf.cast(combine_tensor, tf.bool), combine_tensor.dtype)
return dispatch_tensor, combine_tensor, loss
def call(self, context, x, losses=None):
"""Call the layer."""
params = self.make_params(context)
q = params.compute_q(x)
if self.shared_kv:
kv = params.compute_kv(x)
k = kv
v = kv
else:
k = params.compute_k(x)
v = params.compute_v(x)
if context.mode == "incremental":
if self.shared_kv:
prev_kv, = context.get_states(1)
else:
prev_k, prev_v = context.get_states(2)
current_position = mtf.equal(
mtf.range(context.mesh, self.window_dim, dtype=tf.int32),
mtf.mod(context.position, self.radius))
if self.shared_kv:
kv = mtf.where(current_position,
kv,
prev_kv,
output_shape=prev_kv.shape)
k = kv
v = kv
context.record_new_states([kv])
else:
k = mtf.where(current_position,
params.compute_k(x),
prev_k,
output_shape=prev_k.shape)
v = mtf.where(current_position,
params.compute_v(x),
prev_v,
output_shape=prev_v.shape)
context.record_new_states([k, v])
window_pos = mtf.range(context.mesh, self.window_dim, tf.int32)
visible = mtf.greater_equal(context.position, window_pos)
bias = attention.visibility_mask_to_attention_bias(
visible, context.activation_dtype)
o = attention.attention(
q, k, v, self.window_dim, self.kv_dim, self.kv_dim, bias,
**self.attention_kwargs_from_context(context))
elif context.length_dim.size <= max(256, self.radius * 4):
# nothing fancy - just do full attention and mask
memory_length = self.rename_length_to_memory_length(
context.position, context)
o = attention.attention(
q, self.rename_length_to_memory_length(k, context),
self.rename_length_to_memory_length(v, context),
self.memory_length(context), self.kv_dim, self.kv_dim,
self.compute_bias(context, memory_length, x),
**self.attention_kwargs_from_context(context))
else:
# fancy local attention algorithm
o = attention.local_attention_1d(
q=q,
k=k,
v=None if self.shared_kv else v,
length_dim=context.length_dim,
key_dim=self.kv_dim,
value_dim=self.kv_dim,
length_dim_num_splits=1, # TODO(noam): look at the layout
autoregressive=context.model.fully_autoregressive,
radius=self.radius,
sequence_id=context.sequence_id,
write_priority=context.write_priority,
read_priority=context.read_priority,
attention_kwargs=self.attention_kwargs_from_context(context))
if context.mode == "first_part":
window_pos = mtf.range(context.mesh, self.window_dim, tf.int32)
pos = mtf.range(context.mesh, context.length_dim, tf.int32)
select_recent = mtf.cast(
mtf.equal(mtf.mod(pos, self.radius), window_pos), x.dtype)
select_recent *= mtf.cast(mtf.less(pos, context.initial_position),
x.dtype)
select_recent *= mtf.cast(
mtf.greater_equal(pos, context.initial_position - self.radius),
x.dtype)
state_shape = (k.shape - [context.length_dim, self.kv_dim] +
[self.window_dim, self.kv_dim])
k_state = mtf.einsum([k, select_recent],
output_shape=state_shape,
reduced_dims=[context.length_dim])
context.new_states.append(k_state)
if not self.shared_kv:
v_state = mtf.einsum([v, select_recent],
output_shape=state_shape,
reduced_dims=[context.length_dim])
context.new_states.append(v_state)
return params.compute_output(o, output_shape=x.shape)
def ut_function(state, step, halting_probability, remainders,
n_updates, previous_state):
"""implements act (position-wise halting).
Args:
state: 3-D Tensor: [batch_size, length, channel]
step: indicates number of steps taken so far
halting_probability: halting probability
remainders: act remainders
n_updates: act n_updates
previous_state: previous state
Returns:
transformed_state: transformed state
step: step+1
halting_probability: halting probability
remainders: act remainders
n_updates: act n_updates
new_state: new state
"""
state = self.step_preprocess(context, state, step)
if self.act_type == "random":
# random as halting probability
p = mtf.random_uniform(context.mesh,
shape=halting_probability.shape.dims,
dtype=context.variable_dtype)
else:
last_dim_name = state.shape.dimension_names[-1]
new_dims = [mtf.Dimension(last_dim_name, 1)]
with tf.variable_scope("sigmoid_activation_for_pondering",
reuse=tf.AUTO_REUSE):
p = mtf.layers.dense(state,
variable_dtype=context.variable_dtype,
reduced_dims=[state.shape.dims[-1]],
new_dims=new_dims,
activation=mtf.sigmoid,
use_bias=True)
if self.act_type == "global":
# average over all positions (as a global halting prob)
p = mtf.reduce_mean(p, reduced_dim=p.shape.dims[1])
p = mtf.squeeze(p)
else:
# maintain position-wise probabilities
new_shape = p.shape.dims[:-1]
p = mtf.reshape(p, new_shape)
# Mask for inputs which have not halted yet
still_running = mtf.cast(mtf.less(halting_probability, 1.0),
context.activation_dtype)
# Mask of inputs which halted at this step
new_halted = mtf.cast(
mtf.greater(halting_probability + p * still_running,
threshold),
context.activation_dtype) * still_running
# Mask of inputs which haven't halted, and didn't halt this step
still_running = mtf.cast(
mtf.less_equal(halting_probability + p * still_running,
threshold),
context.activation_dtype) * still_running
# Add the halting probability for this step to the halting
# probabilities for those input which haven't halted yet
halting_probability += p * still_running
# Compute remainders for the inputs which halted at this step
remainders += new_halted * (1 - halting_probability)
# Add the remainders to those inputs which halted at this step
halting_probability += new_halted * remainders
# Increment n_updates for all inputs which are still running
n_updates += still_running + new_halted
# Compute the weight to be applied to the new state and output
# 0 when the input has already halted
# p when the input hasn't halted yet
# the remainders when it halted this step
input_tensor = p * still_running + new_halted * remainders
update_weights = input_tensor
# apply transformation on the state
transformed_state = state
for _ in range(self.num_inrecurrence_layers):
transformed_state = self.vanilla_transformer_layer(
context, transformed_state, mask)
# update running part in the weighted state and keep the rest
new_state = ((transformed_state * update_weights) +
(previous_state * (1 - update_weights)))
if self.act_type == "accumulated":
# Add in the weighted state
new_state = (transformed_state *
update_weights) + previous_state
step += 1
return (transformed_state, step, halting_probability, remainders,
n_updates, new_state)
