def decode(self, input_batch: NestedMap) -> Tuple[NestedMap, NestedMap]:
"""Greedy decodes the input_batch.
Args:
input_batch: The input batch, with fields like `.ids`.
Returns:
- metrics, a NestedMap containing str keys and (metrics, weight) pairs.
- A NestedMap like `input_batch`, with `.prefix_lengths` (vector of
randomly generated ints indicating the lengths of prefixes for each
row), and `.output_ids` (matrix of int ids with the decoded output).
"""
p = self.params
if p.decoder.seqlen <= 0:
raise ValueError('Must set p.decoder.seqlen > 0, current value = '
f'{p.decoder.seqlen}')
batch_size = input_batch.ids.shape[0]
maxval = jnp.sum(1 - input_batch.paddings, axis=1).astype(jnp.int32)
minval = jnp.minimum(maxval, p.decoder.min_prefix_len)
prefix_lengths = jax.random.randint(base_layer.next_prng_key(),
[batch_size], minval, maxval + 1,
input_batch.ids.dtype)
decoder_state = self.lm.init_states(
target_batch_size=batch_size,
target_max_length=p.decoder.seqlen)
global_step = base_layer.cur_global_step()
lm_theta = self.lm.local_theta()
def extend_step_fn(states, ids):
with base_layer.JaxContext.new_context(
prng_key=base_layer.next_prng_key(),
global_step=global_step) as jax_context:
jax_context.bind(self.lm, self.lm.vars_to_flax_vars(lm_theta),
[base_layer.SCOPE_AUX_LOSS])
new_states, xent = self.lm.extend_step(states, ids)
return new_states, xent.logits
result = greedy_decode(
extend_step_fn,
decoder_state,
input_batch.ids,
input_batch.paddings,
p.decoder.seqlen,
max_decode_steps=p.decoder.max_decode_steps,
prefix_lengths=prefix_lengths,
eos_id=p.decoder.eos_id)
result.update(input_batch)
metrics = NestedMap(
num_decoded=(jnp.array(0.0, jnp.float32),
jnp.array(batch_size, jnp.float32)))
return metrics, result
def decode(self, input_batch: NestedMap) -> Tuple[NestedMap, NestedMap]:
"""Decodes input_batch.
Args:
input_batch: The input batch, with a field `.src` and `.tgt` corresponding
to source and target, which itself contains the `.ids` and `.paddings.`
Returns:
- metrics, a nestedmap of metrics.
- results, a NestedMap like `input_batch`, with `.output_ids` (matrix of
int ids with the decoded output) as well as the decoded length.
"""
p = self.params
model_theta = self.model.local_theta()
if p.decoder.seqlen <= 0:
raise ValueError('Must set p.decoder.seqlen > 0, current value = '
f'{p.decoder.seqlen}')
batch_size = input_batch.tgt.ids.shape[0]
decoder_state = self.model.init_states(
inputs=input_batch.src.ids,
input_paddings=input_batch.src.paddings,
target_batch_size=batch_size,
target_max_length=p.decoder.seqlen)
global_step = base_layer.cur_global_step()
def extend_step_fn(states, ids):
with base_layer.JaxContext.new_context(
prng_key=base_layer.next_prng_key(),
global_step=global_step) as jax_context:
jax_context.bind(self.model, self.model.vars_to_flax_vars(model_theta),
[base_layer.SCOPE_AUX_LOSS])
new_states, xent = self.model.extend_step(states, ids)
return new_states, xent.logits
result = greedy_decode(
extend_step_fn,
decoder_state,
input_batch.tgt.ids,
input_batch.tgt.paddings,
p.decoder.seqlen,
eos_id=p.decoder.eos_id)
# Prefix lengths are not needed for sequence model decoding.
del result.prefix_lengths
result.update(input_batch)
metrics = NestedMap(
num_decoded=(jnp.array(0.0, jnp.float32),
jnp.array(batch_size, jnp.float32)))
return metrics, result
def body_fprop(self, per_stage_inputs: JTensor, *per_stage_args,
**per_stage_kwargs) -> NestedJTensor:
"""Runs the fprop function of the stages."""
p = self.params
if p.mesh_axis_names is not None:
def annotate(x):
unconstrained_dims = list(range(1, x.ndim))
dims_mapping = (p.weight_split_dims_mapping.stages + [None] *
(x.ndim - 1))
return base_layer.maybe_shard(x, dims_mapping,
p.mesh_axis_names,
unconstrained_dims)
per_stage_inputs = jax.tree_map(annotate, per_stage_inputs)
per_stage_args = jax.tree_map(annotate, per_stage_args)
per_stage_kwargs = jax.tree_map(annotate, per_stage_kwargs)
prng_key = base_layer.next_prng_key()
global_step = base_layer.cur_global_step()
# vmap self.body.fprop to get a leading stage dimension to handle per_stage
# inputs and args.
def _wrapped_fn(theta, per_stage_inputs, *per_stage_args,
**per_stage_kwargs):
with base_layer.JaxContext.new_context(
prng_key=prng_key, global_step=global_step) as jax_ctx:
jax_ctx.bind(self.body, self.body.vars_to_flax_vars(theta),
[base_layer.SCOPE_AUX_LOSS])
res = self.body.fprop(per_stage_inputs, *per_stage_args,
**per_stage_kwargs)
summaries = base_layer.all_summaries()
return res, summaries
res, summaries = jax.vmap(_wrapped_fn)(self.body.local_theta(),
per_stage_inputs,
*per_stage_args,
**per_stage_kwargs)
self._forward_summary(summaries)
return res
def recurrent_func(theta: NestedMap, states_0: NestedMap, inputs: NestedMap,
cell_fn: Callable[[NestedMap, NestedMap, NestedMap],
NestedMap]):
"""Computes a recurrent neural net.
Args:
theta: weights. A `.NestedMap`.
states_0: initial state. A `.NestedMap`.
inputs: inputs. A `.NestedMap`.
cell_fn: A python function which computes::
states_1 = cell_fn(theta, states_0, inputs[t, :])
Returns:
`accumulate_state` and the final state.
"""
input_seq_len = inputs.Flatten()[0].shape[0]
def assert_not_none(x):
assert x is not None
tf.nest.map_structure(assert_not_none, states_0)
tf.nest.map_structure(assert_not_none, inputs)
tf.nest.map_structure(assert_not_none, theta)
def new_cum_state(x):
x1 = jnp.expand_dims(x, 0)
# +1 so that we can store initial_states at position 0.
return jnp.tile(x1, [input_seq_len + 1] + [1] * x.ndim)
cumulative_states = states_0.Transform(new_cum_state)
prng_key = base_layer.next_prng_key()
global_step = base_layer.cur_global_step()
start_time = jnp.array(0, dtype=jnp.uint32)
fwd_initial_loop_vars = NestedMap(cur_time=start_time,
theta=theta,
states_0=states_0,
cumulative_states=cumulative_states,
inputs=inputs)
def same_type_shape(x, y):
assert x.dtype == y.dtype, (x.dtype, y.dtype)
assert x.shape == y.shape, (x.shape, y.shape)
def wrapped_cell_fn(fn_in):
# fn_in is NestedMap containing the following elements:
# - t
# - theta
# - states_0
# - inputs_t
# Start a chain of prng key that also takes into account of time steps.
t = fn_in.t
theta = fn_in.theta
states_0 = fn_in.states_0
inputs_t = fn_in.inputs_t
with base_layer.JaxContext.new_context(prng_key=jax.random.fold_in(
prng_key, t),
global_step=global_step):
# NO side-effect ops are allowed as the enclosing JaxContext is not bound
# to any layer.
states_1 = cell_fn(theta, states_0, inputs_t)
tf.nest.assert_same_structure(states_0, states_1)
tf.nest.map_structure(same_type_shape, states_0, states_1)
return states_1
def wrapped_cell_fn_grad(fn_in, d_fn_out):
# This is roughly the following:
#
# fn_out = wrapped_cell_fn(fn_in)
# d_fn_in = tf.gradient(fn_out, fn_in, d_fn_out)
# return d_fn_in
#
assert isinstance(fn_in, NestedMap)
fn_out, vjp_fn = jax.vjp(wrapped_cell_fn, fn_in)
del fn_out
d_fn_in = vjp_fn(d_fn_out)
assert isinstance(d_fn_in, tuple)
assert len(d_fn_in) == 1
d_fn_in_0 = d_fn_in[0]
# Over-write gradient for t, the time step.
d_fn_in_0.t = jnp.zeros_like(fn_in.t)
tf.nest.assert_same_structure(fn_in, d_fn_in_0)
tf.nest.map_structure(same_type_shape, fn_in, d_fn_in_0)
return d_fn_in_0
def fwd_comp_fn(loop_vars):
# loop_vars is a NestedMap containing the following elements:
# - cur_time
# - theta
# - inputs
# - cumulative_states
# - states_0
t = loop_vars.cur_time
theta = loop_vars.theta
inputs = loop_vars.inputs
cumulative_states = loop_vars.cumulative_states
states_0 = loop_vars.states_0
inputs_t = inputs.Transform(lambda x: x[t])
states_1 = wrapped_cell_fn(
NestedMap(t=t, theta=theta, states_0=states_0, inputs_t=inputs_t))
def set_t(x, x_t):
return x.at[t + 1].set(x_t)
cumulative_states = tf.nest.map_structure(set_t, cumulative_states,
states_1)
loop_out = NestedMap(cur_time=t + 1,
theta=theta,
inputs=inputs,
states_0=states_1,
cumulative_states=cumulative_states)
return loop_out
def fwd_continue_fn(loop_vars):
return loop_vars.cur_time < input_seq_len
# This custom_vjp implementation follows examples here:
# https://jax.readthedocs.io/en/latest/notebooks/Custom_derivative_rules_for_Python_code.html
@jax.custom_vjp
def fwd_loop(loop_vars):
final_loop_vars = jax.lax.while_loop(fwd_continue_fn, fwd_comp_fn,
loop_vars)
return NestedMap(final_states=final_loop_vars.states_0,
cumulative_states=final_loop_vars.cumulative_states)
def loop_fn_vjp_fwd(loop_vars):
loop_fn_out = fwd_loop(loop_vars)
return loop_fn_out, (loop_vars, loop_fn_out.cumulative_states)
def loop_fn_vjp_bwd(res, d_out):
fwd_loop_vars, cumulative_states = res
d_final_states = d_out.final_states
d_cumulative_states = d_out.cumulative_states
start_time = input_seq_len - 1
d_states_1 = tf.nest.map_structure(lambda x, y: x[start_time + 1] + y,
d_cumulative_states, d_final_states)
bwd_loop_vars = NestedMap(
cur_time=start_time,
theta=fwd_loop_vars.theta,
inputs=fwd_loop_vars.inputs,
cumulative_states=cumulative_states,
d_cumulative_states=d_cumulative_states,
d_theta=fwd_loop_vars.theta.Transform(jnp.zeros_like),
d_inputs=fwd_loop_vars.inputs.Transform(jnp.zeros_like),
d_states_1=d_states_1)
def bwd_comp_fn(loop_vars):
t = loop_vars.cur_time
inputs = loop_vars.inputs
inputs_t = inputs.Transform(lambda x: x[t])
states_0 = loop_vars.cumulative_states.Transform(lambda x: x[t])
d_cell_in = wrapped_cell_fn_grad(
NestedMap(t=t,
theta=loop_vars.theta,
states_0=states_0,
inputs_t=inputs_t), loop_vars.d_states_1)
d_theta = tf.nest.map_structure(lambda x, y: x + y,
loop_vars.d_theta, d_cell_in.theta)
d_states_0 = tf.nest.map_structure(lambda x, y: x + y[t],
d_cell_in.states_0,
loop_vars.d_cumulative_states)
def set_t(x, x_t):
return x.at[t].set(x_t)
d_inputs = tf.nest.map_structure(set_t, loop_vars.d_inputs,
d_cell_in.inputs_t)
loop_vars_out = loop_vars.Transform(lambda x: x)
loop_vars_out.d_inputs = d_inputs
loop_vars_out.d_states_1 = d_states_0
loop_vars_out.d_theta = d_theta
loop_vars_out.cur_time = t - 1
return loop_vars_out
def bwd_continue_fn(loop_vars):
return loop_vars.cur_time >= 0
bwd_final_loop_vars = jax.lax.while_loop(bwd_continue_fn, bwd_comp_fn,
bwd_loop_vars)
d_out = fwd_loop_vars.Transform(jnp.zeros_like)
tf.nest.map_structure(same_type_shape, d_out.states_0,
bwd_final_loop_vars.d_states_1)
tf.nest.map_structure(same_type_shape, d_out.theta,
bwd_final_loop_vars.d_theta)
tf.nest.map_structure(same_type_shape, d_out.inputs,
bwd_final_loop_vars.d_inputs)
d_out.states_0 = bwd_final_loop_vars.d_states_1
d_out.theta = bwd_final_loop_vars.d_theta
d_out.inputs = bwd_final_loop_vars.d_inputs
return (d_out, )
fwd_loop.defvjp(loop_fn_vjp_fwd, loop_fn_vjp_bwd)
# Finally, let's simply run the forward loop fn.
fwd_final_loop_vars = fwd_loop(fwd_initial_loop_vars)
fwd_cumulative_states = fwd_final_loop_vars.cumulative_states.Transform(
lambda x: x[1:])
return fwd_final_loop_vars.final_states, fwd_cumulative_states
def scan(carry_init: NestedMap,
xs: NestedMap,
fn: Callable[[NestedMap, NestedMap], Tuple[NestedMap, NestedMap]],
root_layer: Optional[base_layer.BaseLayer] = None,
checkpoint_policy: AutodiffCheckpointType = AutodiffCheckpointType.
SAVE_NOTHING):
"""A simple wrap around jax.lax.scan.
Back-prop is availale through auto-diff.
Args:
carry_init: initial state. A `.NestedMap`.
xs: inputs. A `.NestedMap`. All inputs in time-major.
fn: A python function which computes:
carry, ys[t] = fn(carry, xs[t, :])
root_layer: The root layer within which this jax.lax.scan based while_loop
is carried out. If root_layer is provided, some basic-effort check is
performed to make sure fn is side-effect free. Otherwise, no such checks
are performed.
checkpoint_policy: A AutodiffCheckpointType. How to checkpoint for BProp:
SAVE_NOTHING, SAVE_DOT_ONLY, SAVE_DOT_WITH_NO_BATCH_DIM.
Returns:
(final 'carry', 'ys', stacked summaries).
"""
del root_layer
assert isinstance(carry_init, py_utils.NestedMap)
assert isinstance(xs, py_utils.NestedMap)
# Make a copy of carry_init structure.
carry_init = tf.nest.map_structure(lambda x: x, carry_init)
# "carry" will be augmented with the following three tensors, so make sure
# they don't already exist in the NestedMap.
assert 'time_step' not in carry_init
assert 'prng_key' not in carry_init
assert 'global_step' not in carry_init
def custom_policy(checkpoint_policy: AutodiffCheckpointType):
# TODO(zhangqiaorjc): Configure custom checkpoint policy in expt config
# without introducing enum.
if checkpoint_policy == AutodiffCheckpointType.SAVE_EVERYTHING:
return jax.checkpoint_policies.everything_saveable
if checkpoint_policy == AutodiffCheckpointType.SAVE_DOT_ONLY:
return jax.checkpoint_policies.checkpoint_dots
if checkpoint_policy == AutodiffCheckpointType.SAVE_DOT_WITH_NO_BATCH_DIM:
return jax.checkpoint_policies.checkpoint_dots_with_no_batch_dims
if checkpoint_policy == AutodiffCheckpointType.SAVE_QKV_OUT_PROJ:
return jax.checkpoint_policies.save_only_these_names(
'combined_qkv_proj', 'out_proj')
if checkpoint_policy == AutodiffCheckpointType.SAVE_CONTEXT:
return jax.checkpoint_policies.save_only_these_names('context')
if checkpoint_policy == AutodiffCheckpointType.SAVE_OUT_PROJ:
return jax.checkpoint_policies.save_only_these_names('out_proj')
if checkpoint_policy == AutodiffCheckpointType.SAVE_CONTEXT_AND_OUT_PROJ:
return jax.checkpoint_policies.save_only_these_names(
'context', 'out_proj')
if checkpoint_policy == AutodiffCheckpointType.SAVE_DOT_FOR_MLPERF_200B:
return jax.checkpoint_policies.save_only_these_names(
'combined_qkv_proj', 'query_proj', 'value_proj', 'key_proj',
'context', 'out_proj')
assert checkpoint_policy == AutodiffCheckpointType.SAVE_NOTHING
return jax.checkpoint_policies.nothing_saveable
@functools.partial(ad_checkpoint.checkpoint,
prevent_cse=False,
policy=custom_policy(checkpoint_policy))
def fn_wrap(carry, xs_t):
# carry is augmented with time_step, prng_key, global_step three additional
# tensors to make fn_wrap fully functional.
# Start a new prng_key branch that also depends on the time step.
prng_key_t = jax.random.fold_in(carry.prng_key, carry.time_step)
with base_layer.JaxContext.new_context(prng_key=prng_key_t,
global_step=carry.global_step):
carry_new, ys_t = fn(carry, xs_t)
carry_new.time_step = carry.time_step + 1
# copy over prng_key and global_step
carry_new.prng_key = carry.prng_key
carry_new.global_step = carry.global_step
tf.nest.assert_same_structure(carry_new, carry)
summaries = base_layer.all_summaries()
return carry_new, (ys_t, summaries)
# The initial time step.
time_step = jnp.array(0, dtype=jnp.uint32)
prng_key = base_layer.next_prng_key()
global_step = base_layer.cur_global_step()
carry_init.time_step = time_step
carry_init.prng_key = prng_key
carry_init.global_step = global_step
carry_final, (ys, summaries) = jax.lax.scan(fn_wrap, carry_init, xs)
del carry_final.time_step
del carry_final.global_step
del carry_final.prng_key
return carry_final, ys, summaries
def recurrent_static(theta: NestedMap,
states_0: NestedMap,
inputs: NestedMap,
cell_fn: Callable[[NestedMap, NestedMap, NestedMap],
NestedMap],
root_layer: Optional[base_layer.BaseLayer] = None):
"""A simpler form of Recurrent where num of steps is known statically.
Back-prop is availale through auto-diff.
'padding' in inputs is used to skip certain steps dynamically. If the
'padding' tensor exists, it is expected of a binary 0/1 tensor.
Args:
theta: weights. A `.NestedMap`.
states_0: initial state. A `.NestedMap`.
inputs: inputs. A `.NestedMap`. All inputs in time-major.
cell_fn: A python function which computes::
states_1 = cell_fn(theta, states_0, inputs[t, :])
root_layer: The root layer within which this recurrent_static recurrent loop
is carried out.
Returns:
`accumulate_state` and the final state.
"""
assert 'time_step' not in states_0
# The initial time step.
time_step = jnp.array(0, dtype=jnp.uint32)
# Make a copy of states_0 structure.
states_0 = tf.nest.map_structure(lambda x: x, states_0)
states_0.time_step = time_step
prng_key = base_layer.next_prng_key()
global_step = base_layer.cur_global_step()
# TODO(zhangqiaorjc): Switch to ad_checkpoint.checkpoint after mattjj bug fix.
@jax.checkpoint
def comp_fn(states_0, inputs_t):
# Start a new prng_key branch that also depends on the time step.
if root_layer is not None:
forward_updated_vars_before = tf.nest.map_structure(
lambda x: x, root_layer.forward_updated_vars)
prng_key_t = jax.random.fold_in(prng_key, states_0.time_step)
with base_layer.JaxContext.new_context(prng_key=prng_key_t,
global_step=global_step):
# NO side-effect ops are allowed as the enclosing JaxContext is not bound
# to any layer.
#
# Whether or not we should skip this time step.
if 'padding' in inputs_t:
# We skip if all are padded steps.
skip = jnp.all(inputs_t.padding > 0.5)
else:
skip = jnp.array(False)
def carry_over(args):
states_0, inputs_t = args
del inputs_t
# We simply carry over the states for this time step.
states_1 = tf.nest.map_structure(lambda x: x, states_0)
states_1.time_step = states_0.time_step + 1
return states_1
def do_compute(args):
states_0, inputs_t = args
# Actually carry out the computation.
states_1 = cell_fn(theta, states_0, inputs_t)
states_1.time_step = states_0.time_step + 1
return states_1
if 'padding' in inputs_t:
states_1 = jax.lax.cond(skip, carry_over, do_compute,
(states_0, inputs_t))
else:
states_1 = do_compute((states_0, inputs_t))
tf.nest.assert_same_structure(states_0, states_1)
if root_layer is not None:
forward_updated_vars_after = tf.nest.map_structure(
lambda x: x, root_layer.forward_updated_vars)
def assert_no_change(x, y):
assert (x is None and y is None) or (x is not None
and y is not None)
tf.nest.map_structure(assert_no_change,
forward_updated_vars_before,
forward_updated_vars_after)
return states_1, states_1
final_states, cumulative_states = jax.lax.scan(comp_fn, states_0, inputs)
del final_states.time_step
del cumulative_states.time_step
return final_states, cumulative_states