def test_entmax():
graph = mtf.Graph()
mesh = mtf.Mesh(graph, "my_mesh")
length = mtf.Dimension("tensor_length", 8)
tensor = mtf.range(mesh, length, tf.float32)
output = entmax(tensor)
grad = mtf.gradients([output], [tensor])[0]
sample = sample_categorical(output, length)
mesh_impl = placement_mesh_impl.PlacementMeshImpl(shape=[], layout={}, devices=[""])
lowering = mtf.Lowering(graph, {mesh: mesh_impl})
sample = lowering.export_to_tf_tensor(sample)
grad = lowering.export_to_tf_tensor(grad)
def body_fn(position, ids, *states):
"""One step in the decode loop."""
nonlocal sampling_keep_top_k
context = mtf_transformer.transformer.Context(
model=None,
mesh=inputs.mesh,
batch_dims=batch_dims,
length_dim=length_dim,
variable_dtype=variable_dtype,
mode="incremental",
position=position,
position_is_default=True,
states=states,
new_states=[],
initial_position=position,
sequence_id=None,
encoder_output=encoder_output,
encoder_sequence_id=encoder_sequence_id,
shared_params=shared_params,
encoder_layer_outputs=encoder_layer_outputs,
write_priority=write_priority,
read_priority=read_priority,
inputs=ids,
encoder_inputs=encoder_inputs) if not slow_sampling else None
with tf.variable_scope("gpt2", reuse=tf.AUTO_REUSE):
logits, _, _ = gpt2.model({"inputs": ids},
other_features,
params,
inputs.mesh,
variable_dtype=variable_dtype,
context=context)
if not sampling_use_entmax:
# By default, do top_k sampling of 0.9
if sampling_keep_top_k == -2:
sampling_keep_top_k = int(logits.shape[-1].size * 0.1)
if sampling_keep_top_k != -1:
if sampling_keep_top_k <= 0:
raise ValueError(
"sampling_keep_top_k must either be -1 or positive.")
k_largest = mtf.nth_largest_element(
logits,
n=sampling_keep_top_k,
reduced_dim=other_features["vocab_dim"])
logits = mtf.where(mtf.less_equal(logits, k_largest),
mtf.ones_like(logits) * -1e6, logits)
ids_this_step = mtf.sample_with_temperature(
logits, other_features["vocab_dim"], temperature)
else:
ids_this_step = sample_categorical(entmax(logits))
if slow_sampling:
ids_this_step = mtf.shift(ids_this_step,
offset=1,
dim=length_dim,
wrap=False)
else:
ids_this_step = mtf.reshape(ids_this_step, (batch_dims))
one_hot = mtf.one_hot(position, length_dim, dtype=tf.int32)
one_new_id = ids_this_step * one_hot
new_ids = (1 - one_hot) * ids + one_new_id
new_position = position + 1
ret = [new_position, new_ids]
if context is not None:
ret += context.new_states
return ret