def mapped_update(i, opt_state, batch, state, rng):
"""This is a multi-device version of the update function above."""
# We assume all tensors have the first dimension = n_devices.
weights, slots, opt_params = opt_state
rng, subrng = jax_random.split(rng)
grad_fn = math.grad(model_and_loss_call, has_aux=True)
grads, state = grad_fn(weights, batch, state, rng)
# We do a psum(1.0) here instead of `n_devices` since `n_devices` is just
# the number of devices on this host machine, however psum goes over all
# devices of all hosts (ex: a TPU pod) and we need to be averaging over all
# of them.
grads = jax.tree_util.tree_map(
lambda g: math.psum(g, 'batch') / math.psum(1.0, 'batch'), grads)
return optimizer.tree_update(i, grads, weights, slots,
opt_params), state, subrng
def forward_with_state(self, xs, weights, state, rng):
self._validate_forward_inputs(xs)
(step, layers_state) = state
# Get N+1 rngs, N for running layers and one extra.
rngs = _split_rngs(rng, self._n_layers + 1)
rng0, rngs = rngs[0], rngs[1:]
if not self.sublayers: # No-op: leave args unchanged.
return (xs, (step + 1, layers_state))
# Prepare the stack and do some safety checks as in the parent class.
stack = xs
new_state = []
n_layers = self._n_layers
if n_layers != 1 and len(weights) != n_layers:
raise ValueError(
'number of weights ({}) not equal to number of layers '
'({})'.format(len(weights), n_layers))
if n_layers != 1 and len(layers_state) != n_layers:
raise ValueError(
'length of state ({}) not equal to number of layers '
'({})'.format(len(layers_state), n_layers))
# TODO(chowdhery): try different strategies, also try running not all
# layers backwards by using math.stop_gradient where needed.
# Calculate how many layers to run forward.
if self._mode == 'train':
# warmup goes from 1.0 at start to 0.0 at skipping_warmup_steps and after
w_steps = float(self._skipping_warmup_steps)
warmup = np.maximum(0.0,
(w_steps - step.astype(np.float32)) / w_steps)
# low is the minimum number of layers to *not* skip, from n_layers to 0
low = warmup * float(n_layers)
# high should be so that (high - n_layers) / high = 1.0 - skip_fraction
# because (high - n_layers) / high is the probability we're not skipping
# (after warmup); so high - n_layers = high - high * skip_fraction
high = float(n_layers) / self._skip_fraction
# We want the same rng0 on all cores.
if math.device_count() > 1:
rng0 = math.psum(rng0, 'batch')
n_forward_layers = random.uniform(rng0, (), np.float32, low, high)
else:
n_forward_layers = float(n_layers)
# Run layers skipping after a certain number.
cur_layer_idx = 0.0
for layer, p, s, rng in zip(self.sublayers, weights, layers_state,
rngs):
inputs = _inputs_from_stack(layer, stack)
outputs, s = math.cond( # Skip (do identity) if > n_forward_layers.
pred=(math.lt(cur_layer_idx, n_forward_layers)),
true_operand=(inputs, p, s, rng), # This tuple is t below.
true_fun=(lambda t: layer.pure_fn(t[0], t[1], t[2], t[3])), # pylint: disable=cell-var-from-loop
false_operand=(inputs, p, s, rng),
false_fun=(lambda t: (t[0], t[2])), # return (inputs, state)
)
stack = _outputs_onto_stack(layer, outputs, stack)
new_state.append(s)
cur_layer_idx += 1.0
return stack, (step + 1, new_state)
