def mtr_lm_v1(num_heads=8, num_memory_heads=0):
"""Model incorporating mixture-of-experts, local and global attention.
~6B parameters
32 experts in 3 hierarchichal moe layers.
Args:
num_heads: an optional integer
num_memory_heads: an optional integer
Returns:
a hparams
"""
hparams = mtr_lm_dense(0)
local_att = transformer_layers.LocalSelfAttention(
num_heads=num_heads,
num_memory_heads=num_memory_heads,
key_value_size=128)
att = transformer_layers.SelfAttention(num_heads=num_heads,
num_memory_heads=num_memory_heads,
key_value_size=128)
drd = transformer_layers.DenseReluDense(hidden_size=2048)
hmoe = moe.MoE2D(expert_x=8, expert_y=4, hidden_size=32768)
hparams.layer_stack = transformer.LayerStack(
([local_att, local_att, drd, att, drd, local_att, local_att, hmoe] *
4)[:-1])
hparams.mesh_shape = "b0:4;b1:8"
hparams.layout = "outer_batch:b0;inner_batch:b1,expert_x:b1,expert_y:b0"
hparams.outer_batch_size = 4
return hparams
def layer_stack(include_encdec_attention):
"""Create a layer stack.
Args:
include_encdec_attention: a boolean
Returns:
a LayerStack
"""
ret = []
for _ in xrange(FLAGS.num_layers):
ret.append(
transformer_layers.SelfAttention(
num_heads=FLAGS.num_heads,
key_value_size=FLAGS.d_kv,
attention_kwargs={"dropout_rate": FLAGS.dropout}))
if include_encdec_attention:
ret.append(
transformer_layers.EncDecAttention(
num_heads=FLAGS.num_heads,
key_value_size=FLAGS.d_kv,
attention_kwargs={"dropout_rate": FLAGS.dropout}))
ret.append(
transformer_layers.DenseReluDense(
hidden_size=FLAGS.d_ff,
dropout_rate=FLAGS.dropout))
return transformer.LayerStack(ret)
def self_attention_layer(hparams, prefix):
"""Create self-attention layer based on hyperparameters."""
return transformer_layers.SelfAttention(
num_heads=hparams.get(prefix + "num_heads"),
num_memory_heads=hparams.get(prefix + "num_memory_heads"),
key_value_size=hparams.d_kv,
shared_kv=hparams.get(prefix + "shared_kv", False),
attention_kwargs=attention_kwargs_from_hparams(hparams))
def my_layer_stack(hparams):
return transformer.LayerStack([
transformer_layers.SelfAttention(
num_heads=hparams.num_heads,
key_value_size=hparams.d_kv,
dropout_rate=hparams.attention_dropout),
transformer_layers.DenseReluDense(
hidden_size=hparams.d_ff,
dropout_rate=hparams.layer_prepostprocess_dropout),
] * hparams.num_hidden_layers)
def mtf_unitransformer_all_layers_tiny():
"""Test out all the layers on local CPU."""
hparams = mtf_unitransformer_tiny()
hparams.layer_stack = transformer.LayerStack(
[transformer_layers.SelfAttention(num_heads=4),
transformer_layers.LocalSelfAttention(num_heads=4),
moe.MoE1D(num_experts=4, hidden_size=512),
moe.MoE2D(expert_x=4, expert_y=4, hidden_size=512),
transformer_layers.DenseReluDense(hidden_size=512)])
return hparams
def __init__(self, base_num_heads):
"""Create an DecoderAttentionLayer.
Args:
base_num_heads: a positive integer, the base number of heads the attention
layers are using.
"""
self._self_attention = transformer_layers.SelfAttention(num_heads=2 *
base_num_heads)
self._enc_dec_attention = transformer_layers.EncDecAttention(
num_heads=base_num_heads)
def mtf_transformer2_all_layers_tiny():
"""Test out all the layers on local CPU."""
hparams = mtf_transformer2_base()
hparams.batch_size = 2
hparams.mesh_shape = ""
hparams.d_model = 128
hparams.layer_stack = transformer.LayerStack([
transformer_layers.SelfAttention(num_heads=4),
transformer_layers.LocalSelfAttention(num_heads=4),
moe.MoE1D(num_experts=4, hidden_size=512),
moe.MoE2D(expert_x=4, expert_y=4, hidden_size=512),
transformer_layers.DenseReluDense(hidden_size=512)
])
return hparams
def default_layer_stack_with_encoder_attention(hparams):
return transformer.LayerStack(
[
transformer_layers.SelfAttention(
num_heads=hparams.num_heads,
key_value_size=hparams.d_kv,
dropout_rate=hparams.attention_dropout),
transformer_layers.EncDecAttention(
num_heads=hparams.num_heads,
key_value_size=hparams.d_kv,
dropout_rate=hparams.attention_dropout),
transformer_layers.DenseReluDense(
hidden_size=hparams.d_ff, dropout_rate=hparams.relu_dropout),
] * hparams.num_hidden_layers,
dropout_rate=hparams.layer_prepostprocess_dropout,
norm_epsilon=hparams.norm_epsilon)
def simple_layer_stack(include_encdec_attention,
num_layers=6,
d_ff=2048,
num_heads=8,
d_kv=128,
dropout_rate=0.1):
"""Create a layer stack.
Args:
include_encdec_attention: a boolean
num_layers: an integer
d_ff: an integer
num_heads: an integer
d_kv: an integer
dropout_rate: a float
Returns:
a LayerStack
"""
ret = []
for _ in xrange(num_layers):
ret.append(
transformer_layers.SelfAttention(
num_heads=num_heads,
key_value_size=d_kv,
attention_kwargs={"dropout_rate": dropout_rate}))
if include_encdec_attention:
ret.append(
transformer_layers.EncDecAttention(
num_heads=num_heads,
key_value_size=d_kv,
attention_kwargs={"dropout_rate": dropout_rate}))
ret.append(
transformer_layers.DenseReluDense(hidden_size=d_ff,
dropout_rate=dropout_rate))
return transformer.LayerStack(ret)
def __init__(self, base_num_heads, key_value_size, dropout_rate):
self._self_attention = transformer_layers.SelfAttention(
num_heads=int(2 * base_num_heads),
key_value_size=int(key_value_size / 2),
dropout_rate=dropout_rate)
def create_dummy_model(mesh,
shapes,
n_blocks=2,
block_param_size_str="2_2",
block_repeat_size_str="1_1"):
"""Creates a dummy model and layer stack with 4-dimensional input."""
assert len(shapes) == 4
outer_batch_size, batch_size, length, d_model = shapes
batch_dim = mtf.Dimension("batch", batch_size)
outer_batch_dim = mtf.Dimension("outer_batch", outer_batch_size)
length_dim = mtf.Dimension("length", length)
block_param_size = list(map(int, block_param_size_str.split("_")))
block_repeat_size = list(map(int, block_repeat_size_str.split("_")))
sublayers_initial = [
transformer.sublayer_dropout,
]
sublayers_per_layer = [
transformer.sublayer_rms_norm,
transformer.sublayer_call_layer,
transformer.sublayer_dropout,
transformer.sublayer_residual,
]
sublayers_final = [
transformer.sublayer_rms_norm,
transformer.sublayer_dropout,
]
submodules = [
transformer_layers.SelfAttention(),
transformer_layers.DenseReluDense()
]
n_sublayers = np.array(block_param_size).prod()
layers = submodules * n_sublayers
layer_stack = funnel_transformer.FunnelTransformerLayerStack(
layers=layers,
n_blocks=n_blocks,
block_param_size=block_param_size,
block_repeat_size=block_repeat_size,
sublayers_initial=sublayers_initial,
sublayers_per_layer=sublayers_per_layer,
sublayers_final=sublayers_final)
model = transformer.Unitransformer(input_vocab_size=10,
output_vocab_size=10,
autoregressive=False,
max_length=8,
d_model=d_model,
layer_stack=layer_stack)
context = transformer.Context(model=model,
mesh=mesh,
batch_dims=[batch_dim, outer_batch_dim],
length_dim=length_dim,
variable_dtype=mtf.VariableDType(tf.float32),
sequence_id=mtf.ones(mesh,
mtf.Shape([length_dim
])),
position=mtf.range(mesh,
length_dim,
dtype=tf.int32))
return layer_stack, context
