使用 TensorFlow 和 Amazon SageMaker 在 TB 级数据上创建、训练和部署十亿参数的语言模型

class TokenAndPositionEmbedding(tf.keras.layers.Layer): """ Creates two separate embedding layers: one for tokens and one for token index (positions). """ def __init__(self, maxlen, vocab_size, embed_dim): super(TokenAndPositionEmbedding, self).__init__() self.token_emb = tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=embed_dim) self.pos_emb = tf.keras.layers.Embedding(input_dim=maxlen, output_dim=embed_dim) def call(self, x): maxlen = tf.shape(x)[-1] # positions are represented by a token's index positions = tf.range(start=0, limit=maxlen, delta=1) positions = self.pos_emb(positions) # token embedding x = self.token_emb(x) # return sum as input return x + positions

class TransformerBlock(tf.keras.layers.Layer): def __init__(self, embed_dim, num_heads, ff_dim, rate=0.1): # self attention layer super(TransformerBlock, self).__init__() self.att = tf.keras.layers.MultiHeadAttention( num_heads=num_heads, key_dim=embed_dim) # feed forward layer self.ffn = [tf.keras.layers.Dense(ff_dim, activation="relu"), tf.keras.layers.Dense(embed_dim)] # layer normalization self.layernorm1 = tf.keras.layers.LayerNormalization(epsilon=1e-6) self.layernorm2 = tf.keras.layers.LayerNormalization(epsilon=1e-6) # dropout self.dropout1 = tf.keras.layers.Dropout(rate) self.dropout2 = tf.keras.layers.Dropout(rate) def call(self, inputs): # getting batch size and seq len from input shape input_shape = tf.shape(inputs) batch_size = input_shape[0] seq_len = input_shape[1] # decoder casual mask casual_mask = casual_attention_mask(batch_size, seq_len, seq_len, tf.bool) # self attention forward pass attention_output = self.att(inputs, inputs, attention_mask=causal_mask) # dense layers, dropout and normalization attention_output = self.dropout1(attention_output) ffn_output = self.ffn[0](out1) ffn_output = self.ffn[1](ffn_output) out2 = self.dropout2(ffn_output) return self.layernorm2(out1 + out2)

class MyModel(tf.keras.Model): def __init__(self, maxlen, vocab_size, embed_dim, num_heads, feed_forward_dim, num_layers, learning_rate): super(MyModel, self).__init__(maxlen, vocab_size, embed_dim, num_heads, feed_forward_dim, num_layers, learning_rate) # embedding layer self.embedding_layer = TokenAndPositionEmbedding(maxlen, vocab_size, embed_dim) # transformer blocks self.transformer_blocks = [ TransformerBlock(embed_dim, num_heads, feed_forward_dim) for i in range(num_layers) ] # last dense layer self.dense = tf.keras.layers.Dense(vocab_size) def call(self, inputs, training=None): x_emb = self.embedding_layer(inputs) x = x_emb for transformer_block in self.transformer_blocks: x = transformer_block(x) outputs = self.dense(x) return [outputs, x_emb] def init_train_settings(maxlen, vocab_size, embed_dim, num_heads, feed_forward_dim, num_layers, learning_rate): """ Creates model, optimizer and loss function """ model = MyModel(maxlen, vocab_size, embed_dim, num_heads, feed_forward_dim, num_layers, learning_rate) loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True) optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate) return model, optimizer, loss_fn

import horovod.tensorflow as hvd
# Initialize Horovod
hvd.init() # Pin GPU to be used to process local rank (one GPU per process)
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus: tf.config.experimental.set_memory_growth(gpu, True)
if gpus: tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU') # Build model
... @tf.function
def training_step(texts, labels, first_batch): with tf.GradientTape() as tape: predictions = model(texts, training=True) loss = loss_fn(labels, predictions[0]) # Horovod: add Horovod Distributed GradientTape. tape = hvd.DistributedGradientTape(tape) grads = tape.gradient(loss, model.trainable_variables) opt.apply_gradients(zip(grads, model.trainable_variables)) # Horovod: broadcast initial variable states from rank 0 to all other processes. # This is necessary to ensure consistent initialization of all workers when # training is started with random weights or restored from a checkpoint. # # Note: broadcast should be done after the first gradient step to ensure optimizer # initialization. if first_batch: hvd.broadcast_variables(model.variables, root_rank=0) hvd.broadcast_variables(opt.variables(), root_rank=0) return loss # Horovod: adjust number of steps based on number of GPUs.
for batch, (texts, labels) in enumerate(dataset.take(10000 // hvd.size())): loss = training_step(texts, labels, batch == 0) if batch % 10 == 0 and hvd.local_rank() == 0: print('Step #%dtLoss: %.6f' % (batch, loss)) # Horovod: save checkpoints only on worker 0 to prevent other workers from
# corrupting it.
if hvd.rank() == 0: checkpoint.save(checkpoint_dir)

import smdistributed.modelparallel.tensorflow as smp # SMP: Initialize
smp.init() # SMP: Define smp.DistributedModel the same way as Keras sub-classing API
class MyModel(smp.DistributedModel): def __init__(self, maxlen, vocab_size, embed_dim, num_heads, feed_forward_dim, num_layers, learning_rate): super(MyModel, self).__init__(maxlen, vocab_size, embed_dim, num_heads, feed_forward_dim, num_layers, learning_rate) self.embedding_layer = gpt_model.TokenAndPositionEmbedding(maxlen, vocab_size, embed_dim) self.transformer_blocks = [ gpt_model.TransformerBlock(embed_dim, num_heads, feed_forward_dim) for i in range(num_layers) ] self.dense = tf.keras.layers.Dense(vocab_size) def call(self, inputs, training=None): x_emb = self.embedding_layer(inputs) x = x_emb for transformer_block in self.transformer_blocks: x = transformer_block(x) outputs = self.dense(x) return [outputs, x_emb] # SMP: Define smp.step. Return any tensors needed outside
@smp.step
def get_grads(texts, labels): predictions = model(texts, training=True) loss = loss_fn(labels, predictions[0]) grads = optimizer.get_gradients(loss, model.trainable_variables) return grads, loss, predictions[0] @tf.function
def train_step(texts, labels, first_batch): gradients, loss, predictions = get_grads(texts, labels) # SMP: Accumulate the gradients across microbatches gradients = [g.accumulate() for g in gradients] optimizer.apply_gradients(zip(gradients, model.trainable_variables)) # SMP: Average the loss across microbatches train_loss(loss.reduce_mean()) # SMP: Merge predictions across microbatches train_accuracy(labels, predictions.merge()) return loss.reduce_mean() histories = [] for _ in range(epochs): train_loss.reset_states() train_accuracy.reset_states() for texts, labels in text_ds: for i in range(128): text = tf.expand_dims(texts[0][i], axis=0) label = tf.expand_dims(labels[0][i], axis=0) train_step(text, label) 

from smdebug.profiler.analysis.notebook_utils.training_job import TrainingJob
tj = TrainingJob('SMD-MP-demo-2022-01-21-06-43-23-841', "us-east-1")
tj.wait_for_sys_profiling_data_to_be_available()
system_metrics_reader = tj.get_systems_metrics_reader()

from smdebug.profiler.analysis.notebook_utils.timeline_charts import TimelineCharts view_timeline_charts = TimelineCharts( system_metrics_reader, framework_metrics_reader, select_dimensions=["CPU", "GPU", "I/O"], select_events=["total"], show_workers=False )