Skip to content

GitLab

"docs/references/k8s.md" did not exist on "bb121214c2daaff0f82483ee83ebc8ad636ef7ff"
Commit 3fc035d7 authored by Mohammad Shoeybi's avatar Mohammad Shoeybi
Browse files

Merge branch 'pipeline_refactor' into 'main' 

Pipeline refactor

See merge request ADLR/megatron-lm!254
parents f2d64c00 e270f68a
Show whitespace changes
Inline Side-by-side
Showing with 479 additions and 837 deletions
megatron/model/__init__.py
View file @ 3fc035d7
......@@ -15,16 +15,8 @@
from .fused_layer_norm import MixedFusedLayerNorm as LayerNorm
from .distributed import *
from .bert_model import (BertModel,
BertModelFirstStage,
BertModelIntermediateStage,
BertModelLastStage)
from .gpt_model import (GPTModel,
GPTModelFirstStage,
GPTModelIntermediateStage,
GPTModelLastStage)
from .distributed import DistributedDataParallel
from .bert_model import BertModel
from .gpt_model import GPTModel
from .language_model import get_language_model
from .module import Float16Module
megatron/model/bert_model.py
View file @ 3fc035d7
......@@ -121,17 +121,23 @@ def post_language_model_processing(lm_output, pooled_output,
return lm_loss, binary_logits
class BertModelBase(MegatronModule):
class BertModel(MegatronModule):
"""Bert Language model."""
def __init__(self, num_tokentypes=2, add_binary_head=True,
parallel_output=True):
super(BertModelBase, self).__init__()
def __init__(self,
num_tokentypes=2,
add_binary_head=True,
parallel_output=True,
pre_process=True,
post_process=True):
super(BertModel, self).__init__()
args = get_args()
self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
self.add_binary_head = add_binary_head
self.parallel_output = parallel_output
self.pre_process = pre_process
self.post_process = post_process
init_method = init_method_normal(args.init_method_std)
scaled_init_method = scaled_init_method_normal(args.init_method_std,
......@@ -142,10 +148,12 @@ class BertModelBase(MegatronModule):
add_pooler=self.add_binary_head,
encoder_attn_mask_type=AttnMaskType.padding,
init_method=init_method,
scaled_init_method=scaled_init_method)
scaled_init_method=scaled_init_method,
pre_process=self.pre_process,
post_process=self.post_process)
self.initialize_word_embeddings(init_method_normal)
if mpu.is_pipeline_last_stage():
if self.post_process:
self.lm_head = BertLMHead(
self.word_embeddings_weight().size(0),
args.hidden_size, init_method, args.layernorm_epsilon, parallel_output)
......@@ -156,26 +164,30 @@ class BertModelBase(MegatronModule):
init_method)
self._binary_head_key = 'binary_head'
def set_input_tensor(self, input_tensor):
"""See megatron.model.transformer.set_input_tensor()"""
self.language_model.set_input_tensor(input_tensor)
def forward(self, bert_model_input, attention_mask,
tokentype_ids=None, lm_labels=None):
extended_attention_mask = bert_extended_attention_mask(attention_mask)
kwargs = {}
if mpu.is_pipeline_first_stage():
input_ids = bert_model_input
position_ids = bert_position_ids(input_ids)
args = [input_ids, position_ids, extended_attention_mask]
kwargs['tokentype_ids'] = tokentype_ids
else:
args = [bert_model_input, extended_attention_mask]
lm_output = self.language_model(*args, **kwargs)
if mpu.is_pipeline_last_stage() and self.add_binary_head:
lm_output = self.language_model(
input_ids,
position_ids,
extended_attention_mask,
tokentype_ids=tokentype_ids
)
if self.post_process and self.add_binary_head:
lm_output, pooled_output = lm_output
else:
pooled_output = None
if mpu.is_pipeline_last_stage():
if self.post_process:
return post_language_model_processing(lm_output, pooled_output,
self.lm_head, self.binary_head,
lm_labels,
......@@ -194,15 +206,15 @@ class BertModelBase(MegatronModule):
state_dict_[self._language_model_key] \
= self.language_model.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
if mpu.is_pipeline_last_stage():
if self.post_process:
state_dict_[self._lm_head_key] \
= self.lm_head.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
if mpu.is_pipeline_last_stage() and self.add_binary_head:
if self.post_process and self.add_binary_head:
state_dict_[self._binary_head_key] \
= self.binary_head.state_dict(destination, prefix, keep_vars)
# Save word_embeddings.
if mpu.is_pipeline_last_stage() and not mpu.is_pipeline_first_stage():
if self.post_process and not self.pre_process:
state_dict_[self._word_embeddings_for_head_key] \
= self.word_embeddings.state_dict(destination, prefix, keep_vars)
return state_dict_
......@@ -212,74 +224,13 @@ class BertModelBase(MegatronModule):
self.language_model.load_state_dict(
state_dict[self._language_model_key], strict=strict)
if mpu.is_pipeline_last_stage():
if self.post_process:
self.lm_head.load_state_dict(
state_dict[self._lm_head_key], strict=strict)
if mpu.is_pipeline_last_stage() and self.add_binary_head:
if self.post_process and self.add_binary_head:
self.binary_head.load_state_dict(
state_dict[self._binary_head_key], strict=strict)
# Load word_embeddings.
if mpu.is_pipeline_last_stage() and not mpu.is_pipeline_first_stage():
if self.post_process and not self.pre_process:
self.word_embeddings.load_state_dict(
state_dict[self._word_embeddings_for_head_key], strict=strict)
class BertModel(BertModelBase):
def __init__(self, num_tokentypes=2, add_binary_head=True,
parallel_output=True):
super(BertModel, self).__init__(
num_tokentypes=num_tokentypes,
add_binary_head=add_binary_head,
parallel_output=parallel_output)
def forward(self, input_ids, attention_mask,
tokentype_ids=None, lm_labels=None):
return super(BertModel, self).forward(
input_ids,
attention_mask,
tokentype_ids=tokentype_ids,
lm_labels=lm_labels)
class BertModelFirstStage(BertModelBase):
def __init__(self, num_tokentypes=2):
super(BertModelFirstStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, input_ids, attention_mask,
tokentype_ids=None):
return super(BertModelFirstStage, self).forward(
input_ids,
attention_mask,
tokentype_ids=tokentype_ids)
class BertModelIntermediateStage(BertModelBase):
def __init__(self, num_tokentypes=2):
super(BertModelIntermediateStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, hidden_state, attention_mask):
return super(BertModelIntermediateStage, self).forward(
hidden_state,
attention_mask)
class BertModelLastStage(BertModelBase):
def __init__(self, num_tokentypes=2, add_binary_head=True,
parallel_output=True):
super(BertModelLastStage, self).__init__(
num_tokentypes=num_tokentypes,
add_binary_head=add_binary_head,
parallel_output=parallel_output)
def forward(self, hidden_state, attention_mask,
lm_labels=None):
return super(BertModelLastStage, self).forward(
hidden_state,
attention_mask,
lm_labels=lm_labels)
megatron/model/classification.py
View file @ 3fc035d7
......@@ -28,13 +28,19 @@ from megatron.model.utils import scaled_init_method_normal
from .module import MegatronModule
class ClassificationBase(MegatronModule):
def __init__(self, num_classes, num_tokentypes=2):
super(ClassificationBase, self).__init__(share_word_embeddings=False)
class Classification(MegatronModule):
def __init__(self,
num_classes,
num_tokentypes=2,
pre_process=True,
post_process=True):
super(Classification, self).__init__(share_word_embeddings=False)
args = get_args()
self.num_classes = num_classes
self.pre_process = pre_process
self.post_process = post_process
init_method = init_method_normal(args.init_method_std)
self.language_model, self._language_model_key = get_language_model(
......@@ -43,31 +49,36 @@ class ClassificationBase(MegatronModule):
encoder_attn_mask_type=AttnMaskType.padding,
init_method=init_method,
scaled_init_method=scaled_init_method_normal(args.init_method_std,
args.num_layers))
args.num_layers),
pre_process=self.pre_process,
post_process=self.post_process)
# Multi-choice head.
if mpu.is_pipeline_last_stage():
if self.post_process:
self.classification_dropout = torch.nn.Dropout(args.hidden_dropout)
self.classification_head = get_linear_layer(args.hidden_size,
self.num_classes,
init_method)
self._classification_head_key = 'classification_head'
def set_input_tensor(self, input_tensor):
"""See megatron.model.transformer.set_input_tensor()"""
self.language_model.set_input_tensor(input_tensor)
def forward(self, model_input, attention_mask, tokentype_ids=None):
extended_attention_mask = bert_extended_attention_mask(attention_mask)
kwargs = {}
if mpu.is_pipeline_first_stage():
input_ids = model_input
position_ids = bert_position_ids(input_ids)
args = [input_ids, position_ids, extended_attention_mask]
kwargs['tokentype_ids'] = tokentype_ids
else:
args = [model_input, extended_attention_mask]
lm_output = self.language_model(*args, **kwargs)
if mpu.is_pipeline_last_stage():
lm_output = self.language_model(
input_ids,
position_ids,
extended_attention_mask,
tokentype_ids=tokentype_ids
)
if self.post_process:
_, pooled_output = lm_output
classification_output = self.classification_dropout(pooled_output)
classification_logits = self.classification_head(classification_output)
......@@ -87,7 +98,7 @@ class ClassificationBase(MegatronModule):
state_dict_[self._language_model_key] \
= self.language_model.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
if mpu.is_pipeline_last_stage():
if self.post_process:
state_dict_[self._classification_head_key] \
= self.classification_head.state_dict(
destination, prefix, keep_vars)
......@@ -98,7 +109,7 @@ class ClassificationBase(MegatronModule):
self.language_model.load_state_dict(
state_dict[self._language_model_key], strict=strict)
if mpu.is_pipeline_last_stage():
if self.post_process:
if self._classification_head_key in state_dict:
self.classification_head.load_state_dict(
state_dict[self._classification_head_key], strict=strict)
......@@ -106,55 +117,3 @@ class ClassificationBase(MegatronModule):
print_rank_last('***WARNING*** could not find {} in the checkpoint, '
'initializing to random'.format(
self._classification_head_key))
class Classification(ClassificationBase):
def __init__(self, num_classes, num_tokentypes=2):
super(Classification, self).__init__(
num_classes, num_tokentypes=num_tokentypes)
def forward(self, input_ids, attention_mask,
tokentype_ids=None):
return super(Classification, self).forward(
input_ids,
attention_mask,
tokentype_ids=tokentype_ids)
class ClassificationFirstStage(ClassificationBase):
def __init__(self, num_classes, num_tokentypes=2):
super(ClassificationFirstStage, self).__init__(
num_classes, num_tokentypes=num_tokentypes)
def forward(self, input_ids, attention_mask,
tokentype_ids=None):
return super(ClassificationFirstStage, self).forward(
input_ids,
attention_mask,
tokentype_ids=tokentype_ids)
class ClassificationIntermediateStage(ClassificationBase):
def __init__(self, num_classes, num_tokentypes=2):
super(ClassificationIntermediateStage, self).__init__(
num_classes, num_tokentypes=num_tokentypes)
def forward(self, hidden_state, attention_mask):
return super(ClassificationIntermediateStage, self).forward(
hidden_state,
attention_mask)
class ClassificationLastStage(ClassificationBase):
def __init__(self, num_classes, num_tokentypes=2):
super(ClassificationLastStage, self).__init__(
num_classes, num_tokentypes=num_tokentypes)
def forward(self, hidden_state, attention_mask):
return super(ClassificationLastStage, self).forward(
hidden_state,
attention_mask)
megatron/model/gpt_model.py
View file @ 3fc035d7
......@@ -57,14 +57,20 @@ def post_language_model_processing(lm_output, labels, logit_weights,
return loss
class GPTModelBase(MegatronModule):
class GPTModel(MegatronModule):
"""GPT-2 Language model."""
def __init__(self, num_tokentypes=0, parallel_output=True):
super(GPTModelBase, self).__init__()
def __init__(self,
num_tokentypes=0,
parallel_output=True,
pre_process=True,
post_process=True):
super(GPTModel, self).__init__()
args = get_args()
self.parallel_output = parallel_output
self.pre_process = pre_process
self.post_process = post_process
self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
self.language_model, self._language_model_key = get_language_model(
......@@ -73,24 +79,28 @@ class GPTModelBase(MegatronModule):
encoder_attn_mask_type=AttnMaskType.causal,
init_method=init_method_normal(args.init_method_std),
scaled_init_method=scaled_init_method_normal(args.init_method_std,
args.num_layers))
args.num_layers),
pre_process=self.pre_process,
post_process=self.post_process)
self.initialize_word_embeddings(init_method_normal)
def forward(self, gpt_model_input, attention_mask, labels=None,
def set_input_tensor(self, input_tensor):
"""See megatron.model.transformer.set_input_tensor()"""
self.language_model.set_input_tensor(input_tensor)
def forward(self, input_ids, position_ids, attention_mask, labels=None,
tokentype_ids=None, layer_past=None, get_key_value=False,
forward_method_parallel_output=None):
kwargs = {'layer_past': layer_past, 'get_key_value': get_key_value}
if mpu.is_pipeline_first_stage():
(input_ids, position_ids) = gpt_model_input
args = [input_ids, position_ids, attention_mask]
kwargs['tokentype_ids'] = tokentype_ids
else:
args = [gpt_model_input, attention_mask]
lm_output = self.language_model(*args, **kwargs)
lm_output = self.language_model(
input_ids,
position_ids,
attention_mask,
layer_past=layer_past,
get_key_value=get_key_value)
if mpu.is_pipeline_last_stage():
if self.post_process:
return post_language_model_processing(
lm_output, labels,
self.word_embeddings_weight(),
......@@ -109,7 +119,7 @@ class GPTModelBase(MegatronModule):
= self.language_model.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
# Save word_embeddings.
if mpu.is_pipeline_last_stage() and not mpu.is_pipeline_first_stage():
if self.post_process and not self.pre_process:
state_dict_[self._word_embeddings_for_head_key] \
= self.word_embeddings.state_dict(destination, prefix, keep_vars)
return state_dict_
......@@ -118,79 +128,9 @@ class GPTModelBase(MegatronModule):
"""Customized load."""
# Load word_embeddings.
if mpu.is_pipeline_last_stage() and not mpu.is_pipeline_first_stage():
if self.post_process and not self.pre_process:
self.word_embeddings.load_state_dict(
state_dict[self._word_embeddings_for_head_key], strict=strict)
if self._language_model_key in state_dict:
state_dict = state_dict[self._language_model_key]
self.language_model.load_state_dict(state_dict, strict=strict)
class GPTModel(GPTModelBase):
def __init__(self, num_tokentypes=0, parallel_output=True):
super(GPTModel, self).__init__(
num_tokentypes=num_tokentypes,
parallel_output=parallel_output)
def forward(self, input_ids, position_ids, attention_mask, labels=None,
tokentype_ids=None, layer_past=None, get_key_value=False,
forward_method_parallel_output=None):
return super(GPTModel, self).forward(
(input_ids, position_ids),
attention_mask,
labels=labels,
tokentype_ids=tokentype_ids,
layer_past=layer_past,
get_key_value=get_key_value,
forward_method_parallel_output=forward_method_parallel_output)
class GPTModelFirstStage(GPTModelBase):
def __init__(self, num_tokentypes=0):
super(GPTModelFirstStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, input_ids, position_ids, attention_mask,
tokentype_ids=None, layer_past=None, get_key_value=False):
return super(GPTModelFirstStage, self).forward(
(input_ids, position_ids),
attention_mask,
tokentype_ids=tokentype_ids,
layer_past=layer_past,
get_key_value=get_key_value)
class GPTModelIntermediateStage(GPTModelBase):
def __init__(self, num_tokentypes=0):
super(GPTModelIntermediateStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, hidden_state, attention_mask,
layer_past=None, get_key_value=False):
return super(GPTModelIntermediateStage, self).forward(
hidden_state,
attention_mask,
layer_past=layer_past,
get_key_value=get_key_value)
class GPTModelLastStage(GPTModelBase):
def __init__(self, num_tokentypes=0, parallel_output=True):
super(GPTModelLastStage, self).__init__(
num_tokentypes=num_tokentypes,
parallel_output=parallel_output)
def forward(self, hidden_state, attention_mask, labels=None,
layer_past=None, get_key_value=False,
forward_method_parallel_output=None):
return super(GPTModelLastStage, self).forward(
hidden_state,
attention_mask,
labels=labels,
layer_past=layer_past,
get_key_value=get_key_value,
forward_method_parallel_output=forward_method_parallel_output)
megatron/model/language_model.py
View file @ 3fc035d7
......@@ -46,7 +46,8 @@ def parallel_lm_logits(input_, word_embeddings_weight, parallel_output,
def get_language_model(num_tokentypes, add_pooler,
encoder_attn_mask_type, init_method=None,
scaled_init_method=None, add_decoder=False,
decoder_attn_mask_type=AttnMaskType.causal):
decoder_attn_mask_type=AttnMaskType.causal,
pre_process=True, post_process=True):
"""Build language model and return along with the key to save."""
args = get_args()
......@@ -58,26 +59,17 @@ def get_language_model(num_tokentypes, add_pooler,
args.num_layers)
# Language model.
args = [init_method, scaled_init_method, encoder_attn_mask_type]
kwargs = {}
cls = None
if mpu.is_pipeline_first_stage() and mpu.is_pipeline_last_stage():
cls = TransformerLanguageModel
kwargs['num_tokentypes'] = num_tokentypes
kwargs['add_decoder'] = add_decoder
kwargs['decoder_attn_mask_type'] = decoder_attn_mask_type
kwargs['add_pooler'] = add_pooler
elif mpu.is_pipeline_first_stage() and not mpu.is_pipeline_last_stage():
cls = TransformerLanguageModelFirstStage
kwargs['num_tokentypes'] = num_tokentypes
elif not mpu.is_pipeline_first_stage() and mpu.is_pipeline_last_stage():
cls = TransformerLanguageModelLastStage
kwargs['add_pooler'] = add_pooler
else:
cls = TransformerLanguageModelIntermediateStage
# Language model.
language_model = cls(*args, **kwargs)
language_model = TransformerLanguageModel(
init_method,
scaled_init_method,
encoder_attn_mask_type,
num_tokentypes=num_tokentypes,
add_decoder=add_decoder,
decoder_attn_mask_type=decoder_attn_mask_type,
add_pooler=add_pooler,
pre_process=pre_process,
post_process=post_process
)
# key used for checkpoints.
language_model_key = 'language_model'
......@@ -263,7 +255,7 @@ class Embedding(MegatronModule):
'checkpoint but could not find it', flush=True)
class TransformerLanguageModelBase(MegatronModule):
class TransformerLanguageModel(MegatronModule):
"""Transformer language model.
Arguments:
......@@ -283,10 +275,14 @@ class TransformerLanguageModelBase(MegatronModule):
num_tokentypes=0,
add_decoder=False,
decoder_attn_mask_type=AttnMaskType.causal,
add_pooler=False):
super(TransformerLanguageModelBase, self).__init__()
add_pooler=False,
pre_process=True,
post_process=True):
super(TransformerLanguageModel, self).__init__()
args = get_args()
self.pre_process = pre_process
self.post_process = post_process
self.hidden_size = args.hidden_size
self.num_tokentypes = num_tokentypes
self.init_method = init_method
......@@ -296,7 +292,7 @@ class TransformerLanguageModelBase(MegatronModule):
self.add_pooler = add_pooler
# Embeddings.
if mpu.is_pipeline_first_stage():
if self.pre_process:
self.embedding = Embedding(self.hidden_size,
args.padded_vocab_size,
args.max_position_embeddings,
......@@ -309,7 +305,10 @@ class TransformerLanguageModelBase(MegatronModule):
self.encoder = ParallelTransformer(
self.init_method,
output_layer_init_method,
self_attn_mask_type=self.encoder_attn_mask_type)
self_attn_mask_type=self.encoder_attn_mask_type,
pre_process=self.pre_process,
post_process=self.post_process
)
self._encoder_key = 'encoder'
# Decoder
......@@ -323,26 +322,29 @@ class TransformerLanguageModelBase(MegatronModule):
self_attn_mask_type=self.decoder_attn_mask_type)
self._decoder_key = 'decoder'
if mpu.is_pipeline_last_stage():
if self.post_process:
# Pooler.
if self.add_pooler:
self.pooler = Pooler(self.hidden_size, self.init_method)
self._pooler_key = 'pooler'
def forward(self, enc_language_model_input, enc_attn_mask,
dec_language_model_input=None, dec_attn_mask=None,
def set_input_tensor(self, input_tensor):
""" See megatron.model.transformer.set_input_tensor()"""
self.encoder.set_input_tensor(input_tensor)
def forward(self, enc_input_ids, enc_position_ids, enc_attn_mask,
dec_input_ids=None, dec_position_ids=None, dec_attn_mask=None,
enc_dec_attn_mask=None, tokentype_ids=None, layer_past=None,
get_key_value=False, pooling_sequence_index=0,
enc_hidden_states=None, output_enc_hidden=False):
# Embeddings.
if mpu.is_pipeline_first_stage():
(input_ids, position_ids) = enc_language_model_input
embedding_output = self.embedding(input_ids, position_ids,
if self.pre_process:
embedding_output = self.embedding(enc_input_ids, enc_position_ids,
tokentype_ids=tokentype_ids)
encoder_input = embedding_output
else:
encoder_input = enc_language_model_input
encoder_input = None
# encoder.
if enc_hidden_states is None:
......@@ -353,7 +355,7 @@ class TransformerLanguageModelBase(MegatronModule):
else:
encoder_output = enc_hidden_states.to(encoder_input.dtype)
if mpu.is_pipeline_last_stage():
if self.post_process:
if self.add_pooler:
pooled_output = self.pooler(encoder_output,
pooling_sequence_index)
......@@ -362,13 +364,12 @@ class TransformerLanguageModelBase(MegatronModule):
# output. For example, it is helpful to compute
# similarity between two sequences by average pooling
if not self.add_decoder or output_enc_hidden:
if self.add_pooler and mpu.is_pipeline_last_stage():
if self.add_pooler and self.post_process:
return encoder_output, pooled_output
else:
return encoder_output
# Decoder Embedding
(dec_input_ids, dec_position_ids) = dec_language_model_input
dec_embedding_output = self.embedding(dec_input_ids,
dec_position_ids)
# decoder
......@@ -379,7 +380,7 @@ class TransformerLanguageModelBase(MegatronModule):
encoder_output=encoder_output,
enc_dec_attn_mask=enc_dec_attn_mask)
if self.add_pooler and mpu.is_pipeline_last_stage():
if self.add_pooler and self.post_process:
return decoder_output, encoder_output, pooled_output
else:
return decoder_output, encoder_output
......@@ -389,14 +390,14 @@ class TransformerLanguageModelBase(MegatronModule):
"""For easy load."""
state_dict_ = {}
if mpu.is_pipeline_first_stage():
if self.pre_process:
state_dict_[self._embedding_key] \
= self.embedding.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
state_dict_[self._encoder_key] \
= self.encoder.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
if mpu.is_pipeline_last_stage():
if self.post_process:
if self.add_pooler:
state_dict_[self._pooler_key] \
= self.pooler.state_dict_for_save_checkpoint(
......@@ -412,7 +413,7 @@ class TransformerLanguageModelBase(MegatronModule):
"""Customized load."""
# Embedding.
if mpu.is_pipeline_first_stage():
if self.pre_process:
if self._embedding_key in state_dict:
state_dict_ = state_dict[self._embedding_key]
else:
......@@ -448,7 +449,7 @@ class TransformerLanguageModelBase(MegatronModule):
self.encoder.load_state_dict(state_dict_, strict=strict)
if mpu.is_pipeline_last_stage():
if self.post_process:
# pooler
if self.add_pooler:
assert 'pooler' in state_dict, \
......@@ -461,124 +462,3 @@ class TransformerLanguageModelBase(MegatronModule):
'could not find data for pooler in the checkpoint'
self.decoder.load_state_dict(state_dict[self._decoder_key],
strict=strict)
class TransformerLanguageModel(TransformerLanguageModelBase):
"""Transformer language model (see TransformerLanguageModelBase
for description of arguments).
"""
def __init__(self,
init_method,
output_layer_init_method,
encoder_attn_mask_type,
num_tokentypes=0,
decoder_attn_mask_type=AttnMaskType.causal,
add_decoder=False,
add_pooler=False):
super(TransformerLanguageModel, self).__init__(
init_method,
output_layer_init_method,
encoder_attn_mask_type,
num_tokentypes=num_tokentypes,
add_decoder=add_decoder,
decoder_attn_mask_type=decoder_attn_mask_type,
add_pooler=add_pooler)
def forward(self, enc_input_ids, enc_position_ids, enc_attn_mask,
dec_input_ids=None, dec_position_ids=None, dec_attn_mask=None,
enc_dec_attn_mask=None, tokentype_ids=None, layer_past=None,
get_key_value=False, pooling_sequence_index=0,
enc_hidden_states=None, output_enc_hidden=False):
return super(TransformerLanguageModel, self).forward(
(enc_input_ids, enc_position_ids),
enc_attn_mask,
dec_language_model_input=(dec_input_ids, dec_position_ids),
dec_attn_mask=dec_attn_mask,
enc_dec_attn_mask=enc_dec_attn_mask,
tokentype_ids=tokentype_ids,
layer_past=layer_past,
get_key_value=get_key_value,
pooling_sequence_index=pooling_sequence_index,
enc_hidden_states=enc_hidden_states,
output_enc_hidden=output_enc_hidden
)
class TransformerLanguageModelFirstStage(TransformerLanguageModelBase):
"""Transformer language model, first stage (see
TransformerLanguageModelBase for description of arguments).
"""
def __init__(self,
init_method,
output_layer_init_method,
encoder_attn_mask_type,
num_tokentypes=0):
super(TransformerLanguageModelFirstStage, self).__init__(
init_method,
output_layer_init_method,
encoder_attn_mask_type,
num_tokentypes=num_tokentypes)
def forward(self, input_ids, position_ids, attention_mask,
tokentype_ids=None, layer_past=None, get_key_value=False):
return super(TransformerLanguageModelFirstStage, self).forward(
(input_ids, position_ids),
attention_mask,
tokentype_ids=tokentype_ids,
layer_past=layer_past,
get_key_value=get_key_value
)
class TransformerLanguageModelIntermediateStage(TransformerLanguageModelBase):
"""Transformer language model, intermediate stage (see
TransformerLanguageModelBase for description of arguments).
"""
def __init__(self,
init_method,
output_layer_init_method,
encoder_attn_mask_type):
super(TransformerLanguageModelIntermediateStage, self).__init__(
init_method,
output_layer_init_method,
encoder_attn_mask_type)
def forward(self, hidden_states, attention_mask,
layer_past=None, get_key_value=False):
return super(TransformerLanguageModelIntermediateStage, self).forward(
hidden_states,
attention_mask,
layer_past=layer_past,
get_key_value=get_key_value
)
class TransformerLanguageModelLastStage(TransformerLanguageModelBase):
"""Transformer language model, final stage (see
TransformerLanguageModelBase for description of arguments).
"""
def __init__(self,
init_method,
output_layer_init_method,
encoder_attn_mask_type,
add_pooler=False):
super(TransformerLanguageModelLastStage, self).__init__(
init_method,
output_layer_init_method,
encoder_attn_mask_type,
add_pooler=add_pooler)
def forward(self, hidden_states, attention_mask,
layer_past=None, get_key_value=False,
pooling_sequence_index=0):
return super(TransformerLanguageModelLastStage, self).forward(
hidden_states,
attention_mask,
layer_past=layer_past,
get_key_value=get_key_value,
pooling_sequence_index=pooling_sequence_index,
)
megatron/model/multiple_choice.py
View file @ 3fc035d7
......@@ -28,13 +28,18 @@ from megatron.model.utils import scaled_init_method_normal
from .module import MegatronModule
class MultipleChoiceBase(MegatronModule):
class MultipleChoice(MegatronModule):
def __init__(self, num_tokentypes=2):
super(MultipleChoiceBase, self).__init__(share_word_embeddings=False)
def __init__(self,
num_tokentypes=2,
pre_process=True,
post_process=True):
super(MultipleChoice, self).__init__(share_word_embeddings=False)
args = get_args()
init_method = init_method_normal(args.init_method_std)
self.pre_process = pre_process
self.post_process = post_process
self.language_model, self._language_model_key = get_language_model(
num_tokentypes=num_tokentypes,
......@@ -42,15 +47,21 @@ class MultipleChoiceBase(MegatronModule):
encoder_attn_mask_type=AttnMaskType.padding,
init_method=init_method,
scaled_init_method=scaled_init_method_normal(args.init_method_std,
args.num_layers))
args.num_layers),
pre_process=self.pre_process,
post_process=self.post_process)
# Multi-choice head.
if mpu.is_pipeline_last_stage():
if self.post_process:
self.multichoice_dropout = torch.nn.Dropout(args.hidden_dropout)
self.multichoice_head = get_linear_layer(args.hidden_size, 1,
init_method)
self._multichoice_head_key = 'multichoice_head'
def set_input_tensor(self, input_tensor):
"""See megatron.model.transformer.set_input_tensor()"""
self.language_model.set_input_tensor(input_tensor)
def forward(self, model_input, attention_mask, tokentype_ids=None):
# [batch, choices, sequence] --> [batch * choices, sequence] -->
......@@ -64,22 +75,21 @@ class MultipleChoiceBase(MegatronModule):
attention_mask = attention_mask.view(-1, attention_mask.size(-1))
extended_attention_mask = bert_extended_attention_mask(attention_mask)
kwargs = {}
if mpu.is_pipeline_first_stage():
input_ids = model_input
# Do the same as attention_mask for input_ids, tokentype_ids
assert len(input_ids.shape) == 3
assert len(tokentype_ids.shape) == 3
input_ids = input_ids.view(-1, input_ids.size(-1))
tokentype_ids = tokentype_ids.view(-1, tokentype_ids.size(-1))
position_ids = bert_position_ids(input_ids)
args = [input_ids, position_ids, extended_attention_mask]
kwargs['tokentype_ids'] = tokentype_ids
else:
args = [model_input, extended_attention_mask]
lm_output = self.language_model(*args, **kwargs)
if mpu.is_pipeline_last_stage():
lm_output = self.language_model(
input_ids,
position_ids,
extended_attention_mask,
tokentype_ids=tokentype_ids
)
if self.post_process:
_, pooled_output = lm_output
multichoice_output = self.multichoice_dropout(pooled_output)
multichoice_logits = self.multichoice_head(multichoice_output)
......@@ -99,7 +109,7 @@ class MultipleChoiceBase(MegatronModule):
state_dict_[self._language_model_key] \
= self.language_model.state_dict_for_save_checkpoint(
destination, prefix, keep_vars)
if mpu.is_pipeline_last_stage():
if self.post_process:
state_dict_[self._multichoice_head_key] \
= self.multichoice_head.state_dict(
destination, prefix, keep_vars)
......@@ -110,7 +120,7 @@ class MultipleChoiceBase(MegatronModule):
self.language_model.load_state_dict(
state_dict[self._language_model_key], strict=strict)
if mpu.is_pipeline_last_stage():
if self.post_process:
if self._multichoice_head_key in state_dict:
self.multichoice_head.load_state_dict(
state_dict[self._multichoice_head_key], strict=strict)
......@@ -118,54 +128,3 @@ class MultipleChoiceBase(MegatronModule):
print_rank_last('***WARNING*** could not find {} in the checkpoint, '
'initializing to random'.format(
self._multichoice_head_key))
class MultipleChoice(MultipleChoiceBase):
def __init__(self, num_tokentypes=2):
super(MultipleChoice, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, input_ids, attention_mask,
tokentype_ids=None):
return super(MultipleChoice, self).forward(
input_ids,
attention_mask,
tokentype_ids=tokentype_ids)
class MultipleChoiceFirstStage(MultipleChoiceBase):
def __init__(self, num_tokentypes=2):
super(MultipleChoiceFirstStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, input_ids, attention_mask,
tokentype_ids=None):
return super(MultipleChoiceFirstStage, self).forward(
input_ids,
attention_mask,
tokentype_ids=tokentype_ids)
class MultipleChoiceIntermediateStage(MultipleChoiceBase):
def __init__(self, num_tokentypes=2):
super(MultipleChoiceIntermediateStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, hidden_state, attention_mask):
return super(MultipleChoiceIntermediateStage, self).forward(
hidden_state,
attention_mask)
class MultipleChoiceLastStage(MultipleChoiceBase):
def __init__(self, num_tokentypes=2):
super(MultipleChoiceLastStage, self).__init__(
num_tokentypes=num_tokentypes)
def forward(self, hidden_state, attention_mask):
return super(MultipleChoiceLastStage, self).forward(
hidden_state,
attention_mask)
megatron/model/transformer.py
View file @ 3fc035d7
......@@ -532,12 +532,16 @@ class ParallelTransformer(MegatronModule):
def __init__(self, init_method, output_layer_init_method,
layer_type=LayerType.encoder,
self_attn_mask_type=AttnMaskType.padding):
self_attn_mask_type=AttnMaskType.padding,
pre_process=True, post_process=True):
super(ParallelTransformer, self).__init__()
args = get_args()
self.bf16 = args.bf16
self.fp32_residual_connection = args.fp32_residual_connection
self.pre_process = pre_process
self.post_process = post_process
self.input_tensor = None
# Store activation checkpoiting flag.
self.checkpoint_activations = args.checkpoint_activations
......@@ -577,10 +581,11 @@ class ParallelTransformer(MegatronModule):
else:
# Each stage gets a contiguous set of layers.
offset = mpu.get_pipeline_model_parallel_rank() * self.num_layers
self.layers = torch.nn.ModuleList(
[build_layer(i + 1 + offset) for i in range(self.num_layers)])
if mpu.is_pipeline_last_stage():
if self.post_process:
# Final layer norm before output.
self.final_layernorm = LayerNorm(
args.hidden_size,
......@@ -615,6 +620,16 @@ class ParallelTransformer(MegatronModule):
return hidden_states
def set_input_tensor(self, input_tensor):
"""Set input tensor to be used instead of forward()'s input.
When doing pipeline parallelism the input from the previous
stage comes from communication, not from the input, so the
model's forward_step_func won't have it. This function is thus
used by internal code to bypass the input provided by the
forward_step_func"""
self.input_tensor = input_tensor
def forward(self, hidden_states, attention_mask, layer_past=None,
get_key_value=False, encoder_output=None, enc_dec_attn_mask=None):
......@@ -628,7 +643,7 @@ class ParallelTransformer(MegatronModule):
'get_key_value does not work with ' \
'activation checkpointing'
if mpu.is_pipeline_first_stage():
if self.pre_process:
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
# If the input flag for fp32 residual connection is set, convert for float.
if self.fp32_residual_connection:
......@@ -636,6 +651,9 @@ class ParallelTransformer(MegatronModule):
# Otherwise, leave it as is.
else:
hidden_states = hidden_states.transpose(0, 1).contiguous()
else:
# See set_input_tensor()
hidden_states = self.input_tensor
if encoder_output is not None:
encoder_output = encoder_output.transpose(0, 1).contiguous()
......@@ -664,7 +682,7 @@ class ParallelTransformer(MegatronModule):
presents.append(present)
# Final layer norm.
if mpu.is_pipeline_last_stage():
if self.post_process:
# Reverting data format change [s b h] --> [b s h].
hidden_states = hidden_states.transpose(0, 1).contiguous()
output = self.final_layernorm(hidden_states)
......
megatron/schedules.py
View file @ 3fc035d7
......@@ -22,6 +22,20 @@ from megatron import get_num_microbatches
from megatron import get_timers
from megatron import mpu
from megatron import p2p_communication
from megatron.utils import unwrap_model
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.model import Float16Module
def get_forward_backward_func():
args = get_args()
if mpu.get_pipeline_model_parallel_world_size() > 1:
if args.virtual_pipeline_model_parallel_size is not None:
forward_backward_func = forward_backward_pipelining_with_interleaving
else:
forward_backward_func = forward_backward_pipelining_without_interleaving
else:
forward_backward_func = forward_backward_no_pipelining
return forward_backward_func
def forward_step(forward_step_func, data_iterator, model, input_tensor, losses_reduced):
......@@ -34,8 +48,12 @@ def forward_step(forward_step_func, data_iterator, model, input_tensor, losses_r
timers = get_timers()
timers('forward-compute').start()
output_tensor = forward_step_func(data_iterator, model, input_tensor)
unwrapped_model = unwrap_model(
model, (torchDDP, LocalDDP, Float16Module))
unwrapped_model.set_input_tensor(input_tensor)
output_tensor, loss_func = forward_step_func(data_iterator, model)
if mpu.is_pipeline_last_stage():
output_tensor = loss_func(output_tensor)
loss, loss_reduced = output_tensor
output_tensor = loss / get_num_microbatches()
losses_reduced.append(loss_reduced)
......
megatron/text_generation_utils.py
View file @ 3fc035d7
......@@ -26,9 +26,13 @@ import torch.nn.functional as F
from megatron import get_args
from megatron import get_tokenizer
from megatron import mpu
from megatron.training import communicate
from megatron.utils import get_ltor_masks_and_position_ids
from megatron.utils import get_ltor_masks_and_position_ids, unwrap_model
from megatron.p2p_communication import recv_forward, send_forward
# These are needed to unwrap the model, would be nice to put these in megatron.utils if possible?
from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.model import Float16Module
def get_batch(context_tokens):
"""Generate batch from context tokens."""
......@@ -395,55 +399,28 @@ def forward_step(model, tokens, position_ids, attention_mask, tokentype_ids,
layer_past=None, get_key_value=None,
forward_method_parallel_output=None):
# Hidden size changes when not using recompute, need to tell communicate()
# the correct size
# Hidden size changes when not using recompute, need to tell p2p_communicate
# functions the correct size
args = get_args()
orig_seq_length = args.seq_length
args.seq_length = tokens.shape[1]
if not mpu.is_pipeline_first_stage():
input_tensor, _ = communicate(
tensor_send_next=None,
tensor_send_prev=None,
recv_forward=True,
recv_backward=False)
else:
input_tensor = None
input_tensor = recv_forward()
# Forward pass through the model.
if mpu.is_pipeline_first_stage():
assert input_tensor is None
if mpu.is_pipeline_last_stage():
output_tensor = model(tokens, position_ids, attention_mask,
tokentype_ids=tokentype_ids,
layer_past=layer_past,
get_key_value=get_key_value,
forward_method_parallel_output=forward_method_parallel_output)
else:
unwrapped_model = unwrap_model(
model, (torchDDP, LocalDDP, Float16Module))
unwrapped_model.set_input_tensor(input_tensor)
output_tensor = model(tokens, position_ids, attention_mask,
tokentype_ids=tokentype_ids,
layer_past=layer_past,
get_key_value=get_key_value)
elif mpu.is_pipeline_last_stage():
assert input_tensor is not None
output_tensor = model(input_tensor, attention_mask,
layer_past=layer_past,
get_key_value=get_key_value,
forward_method_parallel_output=forward_method_parallel_output)
else:
assert input_tensor is not None
output_tensor = model(input_tensor, attention_mask,
layer_past=layer_past,
get_key_value=get_key_value)
if get_key_value:
output_tensor, layer_past = output_tensor
if not mpu.is_pipeline_last_stage():
communicate(tensor_send_next=output_tensor,
tensor_send_prev=None,
recv_forward=False,
recv_backward=False)
send_forward(output_tensor)
args.seq_length = orig_seq_length
if get_key_value:
......
megatron/training.py
View file @ 3fc035d7
......@@ -196,7 +196,25 @@ def get_model(model_provider_func):
args = get_args()
# Build model on cpu.
model = model_provider_func()
pre_process = mpu.is_pipeline_first_stage()
post_process = mpu.is_pipeline_last_stage()
if mpu.get_pipeline_model_parallel_world_size() > 1 and \
args.virtual_pipeline_model_parallel_size is not None:
model = []
for i in range(args.virtual_pipeline_model_parallel_size):
mpu.set_virtual_pipeline_model_parallel_rank(i)
this_model = model_provider_func(
pre_process=pre_process,
post_process=post_process
)
model.append(this_model)
else:
model = model_provider_func(
pre_process=pre_process,
post_process=post_process
)
if not isinstance(model, list):
model = [model]
......
pretrain_bert.py
View file @ 3fc035d7
......@@ -17,56 +17,30 @@
import torch
import torch.nn.functional as F
from functools import partial
from megatron import get_args
from megatron import print_rank_0
from megatron import get_timers
from megatron import mpu
from megatron.data.dataset_utils import build_train_valid_test_datasets
from megatron.model import (BertModel,
BertModelFirstStage,
BertModelIntermediateStage,
BertModelLastStage)
from megatron.model import BertModel
from megatron.training import pretrain
from megatron.utils import average_losses_across_data_parallel_group
def model_provider():
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
print_rank_0('building BERT model ...')
args = get_args()
num_tokentypes = 2 if args.bert_binary_head else 0
def model_provider_pipelined():
# Determine model based on position of stage in pipeline.
if mpu.is_pipeline_first_stage():
model = BertModelFirstStage(
num_tokentypes=num_tokentypes)
elif mpu.is_pipeline_last_stage():
model = BertModelLastStage(
num_tokentypes=num_tokentypes,
add_binary_head=args.bert_binary_head,
parallel_output=True)
else:
model = BertModelIntermediateStage(
num_tokentypes=num_tokentypes)
return model
args = get_args()
if mpu.get_pipeline_model_parallel_world_size() > 1:
if args.virtual_pipeline_model_parallel_size is not None:
model = []
for i in range(args.virtual_pipeline_model_parallel_size):
mpu.set_virtual_pipeline_model_parallel_rank(i)
model.append(model_provider_pipelined())
else:
model = model_provider_pipelined()
else:
model = BertModel(
num_tokentypes=num_tokentypes,
add_binary_head=args.bert_binary_head,
parallel_output=True)
parallel_output=True,
pre_process=pre_process,
post_process=post_process)
return model
......@@ -96,36 +70,7 @@ def get_batch(data_iterator):
return tokens, types, sentence_order, loss_mask, lm_labels, padding_mask
def forward_step(data_iterator, model, input_tensor):
"""Forward step."""
args = get_args()
timers = get_timers()
# Get the batch.
timers('batch-generator').start()
tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = get_batch(
data_iterator)
timers('batch-generator').stop()
if not args.bert_binary_head:
types = None
# Forward pass through the model.
if mpu.is_pipeline_first_stage():
assert input_tensor is None
if mpu.is_pipeline_last_stage():
output_tensor = model(tokens, padding_mask, tokentype_ids=types,
lm_labels=lm_labels)
else:
output_tensor = model(tokens, padding_mask, tokentype_ids=types)
elif mpu.is_pipeline_last_stage():
assert input_tensor is not None
output_tensor = model(input_tensor, padding_mask, lm_labels=lm_labels)
else:
assert input_tensor is not None
output_tensor = model(input_tensor, padding_mask)
if mpu.is_pipeline_last_stage():
def loss_func(loss_mask, sentence_order, output_tensor):
lm_loss_, sop_logits = output_tensor
lm_loss_ = lm_loss_.float()
......@@ -150,7 +95,26 @@ def forward_step(data_iterator, model, input_tensor):
[lm_loss])
return loss, {'lm loss': averaged_losses[0]}
return output_tensor
def forward_step(data_iterator, model):
"""Forward step."""
args = get_args()
timers = get_timers()
# Get the batch.
timers('batch-generator').start()
tokens, types, sentence_order, loss_mask, lm_labels, padding_mask = get_batch(
data_iterator)
timers('batch-generator').stop()
if not args.bert_binary_head:
types = None
# Forward pass through the model.
output_tensor = model(tokens, padding_mask, tokentype_ids=types,
lm_labels=lm_labels)
return output_tensor, partial(loss_func, loss_mask, sentence_order)
def train_valid_test_datasets_provider(train_val_test_num_samples):
......
pretrain_gpt.py
View file @ 3fc035d7
......@@ -16,50 +16,28 @@
"""Pretrain GPT"""
import torch
from functools import partial
from megatron import get_args
from megatron import print_rank_0
from megatron import get_timers
from megatron import get_tokenizer
from megatron import mpu
from megatron.data.gpt_dataset import build_train_valid_test_datasets
from megatron.model import (GPTModel,
GPTModelFirstStage,
GPTModelIntermediateStage,
GPTModelLastStage)
from megatron.model import GPTModel
from megatron.training import pretrain
from megatron.utils import get_ltor_masks_and_position_ids
from megatron.utils import average_losses_across_data_parallel_group
def model_provider():
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
print_rank_0('building GPT model ...')
def model_provider_pipelined():
# Determine model based on position of stage in pipeline.
if mpu.is_pipeline_first_stage():
model = GPTModelFirstStage(num_tokentypes=0)
elif mpu.is_pipeline_last_stage():
model = GPTModelLastStage(
num_tokentypes=0, parallel_output=True)
else:
model = GPTModelIntermediateStage(
num_tokentypes=0)
return model
args = get_args()
if mpu.get_pipeline_model_parallel_world_size() > 1:
if args.virtual_pipeline_model_parallel_size is not None:
model = []
for i in range(args.virtual_pipeline_model_parallel_size):
mpu.set_virtual_pipeline_model_parallel_rank(i)
model.append(model_provider_pipelined())
else:
model = model_provider_pipelined()
else:
model = GPTModel(num_tokentypes=0, parallel_output=True)
model = GPTModel(
num_tokentypes=0,
parallel_output=True,
pre_process=pre_process,
post_process=post_process
)
return model
......@@ -94,8 +72,18 @@ def get_batch(data_iterator):
return tokens, labels, loss_mask, attention_mask, position_ids
def loss_func(loss_mask, output_tensor):
losses = output_tensor.float()
loss_mask = loss_mask.view(-1).float()
loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()
# Reduce loss for logging.
averaged_loss = average_losses_across_data_parallel_group([loss])
return loss, {'lm loss': averaged_loss[0]}
def forward_step(data_iterator, model, input_tensor):
def forward_step(data_iterator, model):
"""Forward step."""
args = get_args()
timers = get_timers()
......@@ -106,31 +94,10 @@ def forward_step(data_iterator, model, input_tensor):
data_iterator)
timers('batch-generator').stop()
# Forward pass through the model.
if mpu.is_pipeline_first_stage():
assert input_tensor is None
if mpu.is_pipeline_last_stage():
output_tensor = model(tokens, position_ids, attention_mask,
labels=labels)
else:
output_tensor = model(tokens, position_ids, attention_mask)
elif mpu.is_pipeline_last_stage():
assert input_tensor is not None
output_tensor = model(input_tensor, attention_mask, labels=labels)
else:
assert input_tensor is not None
output_tensor = model(input_tensor, attention_mask)
if mpu.is_pipeline_last_stage():
losses = output_tensor.float()
loss_mask = loss_mask.view(-1).float()
loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()
# Reduce loss for logging.
averaged_loss = average_losses_across_data_parallel_group([loss])
return loss, {'lm loss': averaged_loss[0]}
return output_tensor
return output_tensor, partial(loss_func, loss_mask)
def train_valid_test_datasets_provider(train_val_test_num_samples):
......
tasks/eval_utils.py
View file @ 3fc035d7
......@@ -17,13 +17,14 @@
import os
import time
from functools import partial
import torch
from megatron import get_args
from megatron import print_rank_last, is_last_rank
from megatron import mpu
from megatron.training import communicate
from megatron.schedules import get_forward_backward_func
from tasks.finetune_utils import build_data_loader
from tasks.finetune_utils import process_batch
......@@ -38,7 +39,7 @@ def accuracy_func_provider(single_dataset_provider):
for datapath in datapaths:
dataset = single_dataset_provider(datapath)
dataloader = build_data_loader(
dataset, args.micro_batch_size, num_workers=args.num_workers,
dataset, args.orig_micro_batch_size, num_workers=args.num_workers,
drop_last=(mpu.get_data_parallel_world_size() > 1))
dataloaders.append((dataset.dataset_name, dataloader))
......@@ -73,14 +74,66 @@ def accuracy_func_provider(single_dataset_provider):
return metrics_func
def calculate_correct_answers(name, model, dataloader,
epoch, output_predictions):
"""Calculate correct over total answers and return prediction if the
`output_predictions` is true."""
args = get_args()
forward_backward_func = get_forward_backward_func()
start_time = time.time()
model.eval()
saved_batch_size = args.micro_batch_size
for m in model:
m.eval()
saved_micro_batch_size = args.micro_batch_size
saved_global_batch_size = args.global_batch_size
ds = dataloader.dataset
if hasattr(ds, 'sample_multiplier'):
# If our dataset as a sample_multiplier attribute that means
# each "sample" from the dataset actually has multiple samples
# that will collapse into the batch dimension (for example in
# the RACE dataset that has several options), we need to
# account for that when setting the micro batch size.
sample_multiplier = ds.sample_multiplier
else:
sample_multiplier = 1
micro_batch_size_times_data_parallel = args.orig_micro_batch_size * args.data_parallel_size
num_micro_batches = args.orig_global_batch_size // micro_batch_size_times_data_parallel
def loss_func(output_predictions, labels, output_tensor):
logits = output_tensor
loss_dict = {}
# Add output predictions.
if output_predictions:
assert False
loss_dict['softmaxes'] = torch.nn.Softmax(dim=-1)(
logits.float()).data.cpu().numpy().tolist()
loss_dict['labels'] = labels.data.cpu().numpy().tolist()
loss_dict['ids'] = batch['uid'].cpu().numpy().tolist()
# Compute the correct answers.
predicted = torch.argmax(logits, dim=-1)
corrects = (predicted == labels)
# Add to the counters.
loss_dict['total'] = labels.size(0)
loss_dict['correct'] = corrects.sum().item()
return 0, loss_dict
# defined inside to capture output_predictions
def correct_answers_forward_step(batch, model):
try:
batch_ = next(batch)
except BaseException:
batch_ = batch
tokens, types, labels, attention_mask = process_batch(batch_)
# Forward model.
args = get_args()
output_tensor = model(tokens, attention_mask, tokentype_ids=types)
return output_tensor, partial(loss_func, output_predictions, labels)
with torch.no_grad():
# For all the batches in the dataset.
total = 0
......@@ -92,60 +145,30 @@ def calculate_correct_answers(name, model, dataloader,
labels = []
ids = []
for _, batch in enumerate(dataloader):
# Run the model forward.
tokens, types, labels_, attention_mask = process_batch(batch)
# For evaluation only mode we use drop_last = False to get all the
# samples, which means we might not have a full batch, so we
# adjust batch_size here to actual batch size of data
actual_batch_size = len(labels_)
actual_batch_size = len(batch['label'])
# ... applying sample_multiplier if necessary
ds = dataloader.dataset
if hasattr(ds, 'sample_multiplier'):
actual_batch_size *= ds.sample_multiplier
args.micro_batch_size = actual_batch_size
if not mpu.is_pipeline_first_stage():
input_tensor, _ = communicate(
tensor_send_next=None,
tensor_send_prev=None,
recv_forward=True,
recv_backward=False)
else:
input_tensor = None
args.micro_batch_size = actual_batch_size * sample_multiplier
args.global_batch_size = actual_batch_size * sample_multiplier * num_micro_batches
# Forward model.
if mpu.is_pipeline_first_stage():
assert input_tensor is None
output_tensor = model(tokens, attention_mask, tokentype_ids=types)
else:
assert input_tensor is not None
output_tensor = model(input_tensor, attention_mask)
loss_dicts = forward_backward_func(correct_answers_forward_step, batch, model,
optimizer=None, timers=None, forward_only=True)
if mpu.is_pipeline_last_stage():
logits = output_tensor
# Add output predictions.
for loss_dict in loss_dicts:
if output_predictions:
softmaxes.extend(torch.nn.Softmax(dim=-1)(
logits.float()).data.cpu().numpy().tolist())
labels.extend(labels_.data.cpu().numpy().tolist())
ids.extend(batch['uid'].cpu().numpy().tolist())
# Compute the correct answers.
predicted = torch.argmax(logits, dim=-1)
corrects = (predicted == labels_)
# Add to the counters.
total += labels_.size(0)
correct += corrects.sum().item()
else:
communicate(
tensor_send_next=output_tensor,
tensor_send_prev=None,
recv_forward=False,
recv_backward=False)
model.train()
args.micro_batch_size = saved_batch_size
softmaxes.extend(loss_dict['softmaxes'])
labels.extend(loss_dict['labels'])
ids.extend(loss_dict['ids'])
total += loss_dict['total']
correct += loss_dict['correct']
for m in model:
m.train()
args.micro_batch_size = saved_micro_batch_size
args.global_batch_size = saved_global_batch_size
# Reduce.
if mpu.is_pipeline_last_stage():
......
tasks/finetune_utils.py
View file @ 3fc035d7
......@@ -15,6 +15,8 @@
"""Finetune utilities."""
from functools import partial
import torch
from megatron import get_args
......@@ -46,7 +48,20 @@ def process_batch(batch):
return tokens, types, labels, attention_mask
def _cross_entropy_forward_step(batch, model, input_tensor):
def cross_entropy_loss_func(labels, output_tensor):
logits = output_tensor
# Cross-entropy loss.
loss_func = torch.nn.CrossEntropyLoss()
loss = loss_func(logits.contiguous().float(), labels)
# Reduce loss for logging.
averaged_loss = average_losses_across_data_parallel_group([loss])
return loss, {'lm loss': averaged_loss[0]}
def _cross_entropy_forward_step(batch, model):
"""Simple forward step with cross-entropy loss."""
timers = get_timers()
......@@ -60,25 +75,9 @@ def _cross_entropy_forward_step(batch, model, input_tensor):
timers('batch-generator').stop()
# Forward model.
if mpu.is_pipeline_first_stage():
assert input_tensor is None
output_tensor = model(tokens, attention_mask, tokentype_ids=types)
else:
assert input_tensor is not None
output_tensor = model(input_tensor, attention_mask)
if mpu.is_pipeline_last_stage():
logits = output_tensor
# Cross-entropy loss.
loss_func = torch.nn.CrossEntropyLoss()
loss = loss_func(logits.contiguous().float(), labels)
# Reduce loss for logging.
averaged_loss = average_losses_across_data_parallel_group([loss])
return loss, {'lm loss': averaged_loss[0]}
return output_tensor
return output_tensor, partial(cross_entropy_loss_func, labels)
def build_data_loader(dataset, micro_batch_size, num_workers, drop_last):
......@@ -135,7 +134,14 @@ def _build_train_valid_dataloaders(train_dataset, valid_dataset):
# This is necessary so pipeline transfers know what size they are
# and the LR schedule, which is based on samples seen, gets set
# correctly.
args.orig_micro_batch_size = args.micro_batch_size
args.orig_global_batch_size = args.global_batch_size
if hasattr(train_dataset, 'sample_multiplier'):
# If our dataset as a sample_multiplier attribute that means
# each "sample" from the dataset actually has multiple samples
# that will collapse into the batch dimension (for example in
# the RACE dataset that has several options), we need to
# account for that when setting the micro batch size.
args.micro_batch_size *= train_dataset.sample_multiplier
args.global_batch_size *= train_dataset.sample_multiplier
......@@ -149,7 +155,8 @@ def _train(model, optimizer, lr_scheduler, forward_step,
timers = get_timers()
# Turn on training mode which enables dropout.
model.train()
for m in model:
m.train()
# Tracking loss.
losses_dict_sum = {}
......@@ -180,10 +187,8 @@ def _train(model, optimizer, lr_scheduler, forward_step,
start_iteration = 0
# Train for one step.
losses_dict, skipped_iter, grad_norm = train_step(forward_step,
batch, model,
optimizer,
lr_scheduler)
out = train_step(forward_step, batch, model, optimizer, lr_scheduler)
losses_dict, skipped_iter, grad_norm, num_zeros_in_grad = out
iteration += 1
# Logging.
......@@ -195,7 +200,7 @@ def _train(model, optimizer, lr_scheduler, forward_step,
iteration,
optimizer.get_loss_scale().item(),
report_memory_flag, skipped_iter,
grad_norm, params_norm)
grad_norm, params_norm, num_zeros_in_grad)
# Autoresume
if args.adlr_autoresume and \
......@@ -231,6 +236,9 @@ def finetune(train_valid_datasets_provider, model_provider,
args = get_args()
timers = get_timers()
assert args.rampup_batch_size is None, \
'batch size scaling is not supported for finetuning'
# Train and validation data loaders.
timers('train/valid/test dataset/dataloder').start()
if args.epochs > 0:
......
tasks/glue/finetune.py
View file @ 3fc035d7
......@@ -19,7 +19,7 @@ from megatron import get_args
from megatron import print_rank_0
from megatron import get_tokenizer
from megatron import mpu
from megatron.model.classification import Classification, ClassificationFirstStage, ClassificationIntermediateStage, ClassificationLastStage
from megatron.model.classification import Classification
from tasks.eval_utils import accuracy_func_provider
from tasks.finetune_utils import finetune
......@@ -39,25 +39,14 @@ def glue_classification(num_classes, Dataset,
return train_dataset, valid_dataset
def model_provider():
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
args = get_args()
print_rank_0('building classification model for {} ...'.format(
args.task))
if mpu.get_pipeline_model_parallel_world_size() > 1:
# Determine model based on position of stage in pipeline.
if mpu.is_pipeline_first_stage():
model = ClassificationFirstStage(
num_classes=num_classes, num_tokentypes=2)
elif mpu.is_pipeline_last_stage():
model = ClassificationLastStage(
num_classes=num_classes, num_tokentypes=2)
else:
model = ClassificationIntermediateStage(
num_classes=num_classes, num_tokentypes=2)
else:
model = Classification(num_classes=num_classes, num_tokentypes=2)
model = Classification(num_classes=num_classes, num_tokentypes=2,
pre_process=pre_process, post_process=post_process)
return model
......
tasks/main.py
View file @ 3fc035d7
......@@ -70,6 +70,11 @@ if __name__ == '__main__':
initialize_megatron(extra_args_provider=get_tasks_args)
args = get_args()
if args.num_layers_per_virtual_pipeline_stage is not None:
print("Interleaved pipeline schedule is not yet supported for downstream tasks.")
exit()
if args.task == 'RACE':
from race.finetune import main
elif args.task in ['MNLI', 'QQP']:
......
tasks/race/data.py
View file @ 3fc035d7
......@@ -39,6 +39,8 @@ class RaceDataset(Dataset):
print_rank_0(' >> total number of samples: {}'.format(
len(self.samples)))
# This indicates that each "sample" has multiple samples that
# will collapse into batch dimension
self.sample_multiplier = NUM_CHOICES
def __len__(self):
......
tasks/race/finetune.py
View file @ 3fc035d7
......@@ -19,7 +19,7 @@ from megatron import get_args
from megatron import print_rank_0
from megatron import get_tokenizer
from megatron import mpu
from megatron.model.multiple_choice import MultipleChoice, MultipleChoiceFirstStage, MultipleChoiceIntermediateStage, MultipleChoiceLastStage
from megatron.model.multiple_choice import MultipleChoice
from tasks.eval_utils import accuracy_func_provider
from tasks.finetune_utils import finetune
from tasks.race.data import RaceDataset
......@@ -38,20 +38,13 @@ def train_valid_datasets_provider():
return train_dataset, valid_dataset
def model_provider():
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
print_rank_0('building multichoice model for RACE ...')
if mpu.get_pipeline_model_parallel_world_size() > 1:
# Determine model based on position of stage in pipeline.
if mpu.is_pipeline_first_stage():
model = MultipleChoiceFirstStage(num_tokentypes=2)
elif mpu.is_pipeline_last_stage():
model = MultipleChoiceLastStage(num_tokentypes=2)
else:
model = MultipleChoiceIntermediateStage(num_tokentypes=2)
else:
model = MultipleChoice(num_tokentypes=2)
model = MultipleChoice(num_tokentypes=2,
pre_process=pre_process,
post_process=post_process)
return model
......
tasks/zeroshot_gpt/evaluate.py
View file @ 3fc035d7
......@@ -24,19 +24,24 @@ from megatron import print_rank_0, is_last_rank
from megatron import get_tokenizer
from megatron import mpu
from megatron.checkpointing import load_checkpoint
from megatron.model import GPTModel, GPTModelFirstStage, GPTModelLastStage, GPTModelIntermediateStage
from megatron.training import get_model, communicate
from megatron.utils import get_ltor_masks_and_position_ids
from megatron.model import GPTModel
from megatron.training import get_model
from megatron.utils import get_ltor_masks_and_position_ids, unwrap_model
from megatron.p2p_communication import recv_forward, send_forward
from tasks.finetune_utils import build_data_loader
from .datasets import build_dataset
# These are needed to unwrap the model, would be nice to put these in megatron.utils if possible?
from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
from megatron.model import DistributedDataParallel as LocalDDP
from megatron.model import Float16Module
def get_model_provider(eval_metric):
"""Based on evaluation metric set the parallel-output flag and
return the model provider."""
def model_provider():
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
if eval_metric == 'loss':
......@@ -48,17 +53,8 @@ def get_model_provider(eval_metric):
'is not supported.'.format(eval_metric))
print_rank_0('building GPT model ...')
if mpu.get_pipeline_model_parallel_world_size() > 1:
# Determine model based on position of stage in pipeline.
if mpu.is_pipeline_first_stage():
model = GPTModelFirstStage(num_tokentypes=0)
elif mpu.is_pipeline_last_stage():
model = GPTModelLastStage(
parallel_output=parallel_output, num_tokentypes=0)
else:
model = GPTModelIntermediateStage(num_tokentypes=0)
else:
model = GPTModel(num_tokentypes=0, parallel_output=parallel_output)
model = GPTModel(num_tokentypes=0, parallel_output=parallel_output,
pre_process=pre_process, post_process=post_process)
return model
......@@ -97,33 +93,15 @@ def forward_step(batch, model, eval_metric):
args = get_args()
args.micro_batch_size = len(labels)
# Forward model.
if not mpu.is_pipeline_first_stage():
input_tensor, _ = communicate(
tensor_send_next=None,
tensor_send_prev=None,
recv_forward=True,
recv_backward=False)
else:
input_tensor = None
input_tensor = recv_forward()
# Forward pass through the model.
if mpu.is_pipeline_first_stage():
assert input_tensor is None
if mpu.is_pipeline_last_stage():
output = model(tokens, position_ids, attention_mask)
else:
unwrapped_model = unwrap_model(
model, (torchDDP, LocalDDP, Float16Module))
unwrapped_model.set_input_tensor(input_tensor)
output = model(tokens, position_ids, attention_mask)
else:
assert input_tensor is not None
output = model(input_tensor, attention_mask)
if not mpu.is_pipeline_last_stage():
communicate(tensor_send_next=output,
tensor_send_prev=None,
recv_forward=False,
recv_backward=False)
return None
send_forward(output)
if mpu.is_pipeline_last_stage():
# For loss, return the unreduced loss.
......@@ -214,6 +192,10 @@ def main():
"""Main program."""
args = get_args()
if args.num_layers_per_virtual_pipeline_stage is not None:
print("Interleaved pipeline schedule is not yet supported for text generation.")
exit()
if args.task == 'LAMBADA':
eval_metric = 'accuracy'
elif args.task == 'WIKITEXT103':
......@@ -227,6 +209,9 @@ def main():
if args.load is not None:
_ = load_checkpoint(model, None, None)
assert len(model) == 1, "Above condition should have caught this"
model = model[0]
# Data stuff.
dataset = build_dataset(args.task)
dataloader = build_data_loader(dataset, args.micro_batch_size,
......
tools/generate_samples_gpt.py
View file @ 3fc035d7
......@@ -26,33 +26,19 @@ from megatron import get_tokenizer
from megatron import mpu
from megatron.checkpointing import load_checkpoint
from megatron.initialize import initialize_megatron
from megatron.model import (GPTModel,
GPTModelFirstStage,
GPTModelLastStage,
GPTModelIntermediateStage)
from megatron.model import GPTModel
from megatron.training import get_model
from megatron.text_generation_utils import generate_and_write_samples_unconditional
from megatron.text_generation_utils import generate_samples_input_from_file
from megatron.text_generation_utils import generate_samples_interactive
def model_provider():
def model_provider(pre_process=True, post_process=True):
"""Build the model."""
print_rank_0('building GPT model ...')
args = get_args()
if mpu.get_pipeline_model_parallel_world_size() > 1:
# Determine model based on position of stage in pipeline.
if mpu.is_pipeline_first_stage():
model = GPTModelFirstStage(num_tokentypes=0)
elif mpu.is_pipeline_last_stage():
model = GPTModelLastStage(
num_tokentypes=0, parallel_output=False)
else:
model = GPTModelIntermediateStage(
num_tokentypes=0)
else:
model = GPTModel(num_tokentypes=0, parallel_output=False)
model = GPTModel(num_tokentypes=0, parallel_output=False,
pre_process=pre_process, post_process=post_process)
return model
......@@ -96,12 +82,20 @@ def main():
'no_load_rng': True,
'no_load_optim': True})
args = get_args()
if args.num_layers_per_virtual_pipeline_stage is not None:
print("Interleaved pipeline schedule is not yet supported for text generation.")
exit()
# Set up model and load checkpoint.
model = get_model(model_provider)
args = get_args()
if args.load is not None:
_ = load_checkpoint(model, None, None)
assert len(model) == 1, "Above condition should have caught this"
model = model[0]
# Generate samples.
if args.num_samples == 0:
args.micro_batch_size = 1
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment