Merge branch 'readme_update_jan_2021' into 'main'

Readme update + change gpt2 to gpt

See merge request ADLR/megatron-lm!206

examples/finetune_mnli_distributed.sh

@@ -28,11 +28,11 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
                --num-layers 24 \
                --hidden-size 1024 \
                --num-attention-heads 16 \
-               --batch-size 8 \
+               --micro-batch-size 8 \
                --checkpoint-activations \
                --lr 5.0e-5 \
                --lr-decay-style linear \
-               --warmup 0.065 \
+               --lr-warmup-fraction 0.065 \
                --seq-length 512 \
                --max-position-embeddings 512 \
                --save-interval 500000 \

examples/finetune_race_distributed.sh

@@ -28,11 +28,11 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS ./tasks/main.py \
                --num-layers 24 \
                --hidden-size 1024 \
                --num-attention-heads 16 \
-               --batch-size 4 \
+               --micro-batch-size 4 \
                --checkpoint-activations \
                --lr 1.0e-5 \
                --lr-decay-style linear \
-               --warmup 0.06 \
+               --lr-warmup-fraction 0.06 \
                --seq-length 512 \
                --max-position-embeddings 512 \
                --save-interval 100000 \

examples/pretrain_bert.sh

@@ -9,24 +9,24 @@ python pretrain_bert.py \
        --num-layers 24 \
        --hidden-size 1024 \
        --num-attention-heads 16 \
-       --batch-size 4 \
+       --micro-batch-size 4 \
+       --global-batch-size 8 \
        --seq-length 512 \
        --max-position-embeddings 512 \
        --train-iters 2000000 \
+       --lr-decay-iters 990000 \
        --save $CHECKPOINT_PATH \
        --load $CHECKPOINT_PATH \
        --data-path $DATA_PATH \
        --vocab-file bert-vocab.txt \
        --data-impl mmap \
        --split 949,50,1 \
-       --distributed-backend nccl \
        --lr 0.0001 \
        --min-lr 0.00001 \
        --lr-decay-style linear \
-       --lr-decay-iters 990000 \
+       --lr-warmup-fraction .01 \
        --weight-decay 1e-2 \
        --clip-grad 1.0 \
-       --warmup .01 \
        --log-interval 100 \
        --save-interval 10000 \
        --eval-interval 1000 \

examples/pretrain_bert_distributed.sh

@@ -15,11 +15,11 @@ DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $
 
 python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        pretrain_bert.py \
-       --tensor-model-parallel-size 1 \
        --num-layers 24 \
        --hidden-size 1024 \
        --num-attention-heads 16 \
-       --batch-size 4 \
+       --micro-batch-size 4 \
+       --global-batch-size 32 \
        --seq-length 512 \
        --max-position-embeddings 512 \
        --train-iters 1000000 \
@@ -36,7 +36,7 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --lr-decay-iters 990000 \
        --weight-decay 1e-2 \
        --clip-grad 1.0 \
-       --warmup .01 \
+       --lr-warmup-fraction .01 \
        --log-interval 100 \
        --save-interval 10000 \
        --eval-interval 1000 \

examples/pretrain_bert_distributed_with_mp.sh

@@ -20,8 +20,8 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --num-layers 24 \
        --hidden-size 1024 \
        --num-attention-heads 16 \
-       --batch-size 2 \
-       --num-microbatches-in-minibatch 2 \
+       --micro-batch-size 2 \
+       --global-batch-size 16 \
        --seq-length 512 \
        --max-position-embeddings 512 \
        --train-iters 1000000 \
@@ -38,7 +38,7 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --lr-decay-iters 990000 \
        --weight-decay 1e-2 \
        --clip-grad 1.0 \
-       --warmup .01 \
+       --lr-warmup-fraction .01 \
        --log-interval 100 \
        --save-interval 10000 \
        --eval-interval 1000 \

examples/pretrain_gpt2.sh → examples/pretrain_gpt.sh

@@ -9,11 +9,12 @@ DATA_PATH=<Specify path and file prefix>_text_document
 CHECKPOINT_PATH=<Specify path>
 
 
-python pretrain_gpt2.py \
+python pretrain_gpt.py \
        --num-layers 24 \
        --hidden-size 1024 \
        --num-attention-heads 16 \
-       --batch-size 8 \
+       --micro-batch-size 4 \
+       --global-batch-size 8 \
        --seq-length 1024 \
        --max-position-embeddings 1024 \
        --train-iters 500000 \
@@ -31,7 +32,7 @@ python pretrain_gpt2.py \
        --lr-decay-style cosine \
        --weight-decay 1e-2 \
        --clip-grad 1.0 \
-       --warmup .01 \
+       --lr-warmup-fraction .01 \
        --checkpoint-activations \
        --log-interval 100 \
        --save-interval 10000 \

examples/pretrain_gpt3_175B.sh

+#!/bin/bash
+
+
+#SBATCH <SLURM OPTIONS> --nodes=128 --exclusive --ntasks-per-node=8 --job-name=megatron_gpt3_175b
+
+
+DIR=`pwd`
+DATETIME=`date +'date_%y-%m-%d_time_%H-%M-%S'`
+mkdir -p $DIR/logs
+
+
+DATASET_1="<PATH TO THE FIRST DATASET>"
+DATASET_2="<PATH TO THE SECOND DATASET>"
+DATASET_3="<PATH TO THE THIRD DATASET>"
+DATASET="0.2 ${DATASET_1} 0.3 ${DATASET_2} 0.5 ${DATASET_3}"
+
+
+options=" \
+	--tensor-model-parallel-size 8 \
+	--pipeline-model-parallel-size 16 \
+        --num-layers 96 \
+        --hidden-size 12288 \
+        --num-attention-heads 96 \
+        --seq-length 2048 \
+        --max-position-embeddings 2048 \
+	--micro-batch-size 1 \
+	--global-batch-size 1536 \
+	--rampup-batch-size 16 16 5859375 \
+	--train-samples 146484375 \
+       	--lr-decay-samples 126953125 \
+        --lr-warmup-samples 183105 \
+        --lr 6.0e-5 \
+	--min-lr 6.0e-6 \
+        --lr-decay-style cosine \
+        --log-interval 10 \
+        --eval-iters 40 \
+        --eval-interval 1000 \
+	--data-path ${DATASET} \
+	--vocab-file <PATH TO gpt-vocab.json> \
+	--merge-file <PATH TO gpt-merges.txt> \
+	--save-interval 1000 \
+	--save <PATH TO CHECKPOINTS DIRECTORY> \
+	--load <PATH TO CHECKPOINTS DIRECTORY> \
+        --split 98,2,0 \
+        --clip-grad 1.0 \
+	--weight-decay 0.1 \
+	--adam-beta1 0.9 \
+	--adam-beta2 0.95 \
+	--tensorboard-dir <TENSORBOARD DIRECTORY> \
+        --fp16 \
+	--checkpoint-activations "
+
+
+run_cmd="python -u ${DIR}/pretrain_gpt.py $@ ${options}"
+
+
+srun -l \
+     --container-image "nvcr.io/nvidia/pytorch:20.12-py3" \
+     --container-mounts "<DIRECTORIES TO MOUNT>" \
+     --output=$DIR/logs/%x_%j_$DATETIME.log sh -c "${run_cmd}"
+
+
+set +x
+

examples/pretrain_gpt2_distributed.sh → examples/pretrain_gpt_distributed.sh

@@ -16,12 +16,12 @@ CHECKPOINT_PATH=<Specify path>
 DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
 
 python -m torch.distributed.launch $DISTRIBUTED_ARGS \
-       pretrain_gpt2.py \
-       --tensor-model-parallel-size 1 \
+       pretrain_gpt.py \
        --num-layers 24 \
        --hidden-size 1024 \
        --num-attention-heads 16 \
-       --batch-size 8 \
+       --micro-batch-size 8 \
+       --global-batch-size 64 \
        --seq-length 1024 \
        --max-position-embeddings 1024 \
        --train-iters 500000 \
@@ -39,7 +39,7 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --min-lr 1.0e-5 \
        --weight-decay 1e-2 \
        --clip-grad 1.0 \
-       --warmup .01 \
+       --lr-warmup-fraction .01 \
        --checkpoint-activations \
        --log-interval 100 \
        --save-interval 10000 \

examples/pretrain_gpt2_distributed_with_mp.sh → examples/pretrain_gpt_distributed_with_mp.sh

@@ -16,14 +16,14 @@ CHECKPOINT_PATH=<Specify path>
 DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
 
 python -m torch.distributed.launch $DISTRIBUTED_ARGS \
-       pretrain_gpt2.py \
+       pretrain_gpt.py \
        --tensor-model-parallel-size 2 \
        --pipeline-model-parallel-size 2 \
        --num-layers 24 \
        --hidden-size 1024 \
        --num-attention-heads 16 \
-       --batch-size 4 \
-       --num-microbatches-in-minibatch 2 \
+       --micro-batch-size 4 \
+       --global-batch-size 16 \
        --seq-length 1024 \
        --max-position-embeddings 1024 \
        --train-iters 500000 \
@@ -41,7 +41,7 @@ python -m torch.distributed.launch $DISTRIBUTED_ARGS \
        --min-lr 1.0e-5 \
        --weight-decay 1e-2 \
        --clip-grad 1.0 \
-       --warmup .01 \
+       --lr-warmup-fraction .01 \
        --checkpoint-activations \
        --log-interval 100 \
        --save-interval 10000 \

megatron/data/gpt2_dataset.py → megatron/data/gpt_dataset.py

@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""GPT2 style dataset."""
+"""GPT style dataset."""
 
 import os
 import time
@@ -107,7 +107,7 @@ def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
         if splits[index + 1] > splits[index]:
             documents = np.arange(start=splits[index], stop=splits[index + 1],
                                   step=1, dtype=np.int32)
-            dataset = GPT2Dataset(name, data_prefix,
+            dataset = GPTDataset(name, data_prefix,
                                   documents, indexed_dataset,
                                   train_valid_test_num_samples[index],
                                   seq_length, seed)
@@ -136,7 +136,7 @@ def get_indexed_dataset_(data_prefix, data_impl, skip_warmup):
     return indexed_dataset
 
 
-class GPT2Dataset(torch.utils.data.Dataset):
+class GPTDataset(torch.utils.data.Dataset):
 
     def __init__(self, name, data_prefix, documents, indexed_dataset,
                  num_samples, seq_length, seed):

megatron/model/__init__.py

@@ -34,10 +34,11 @@ from .bert_model import (BertModel,
                          BertModelFirstStage,
                          BertModelIntermediateStage,
                          BertModelLastStage)
-from .gpt2_model import (GPT2Model,
-                         GPT2ModelFirstStage,
-                         GPT2ModelIntermediateStage,
-                         GPT2ModelLastStage)
+from .realm_model import ICTBertModel
+from .gpt_model import (GPTModel,
+                        GPTModelFirstStage,
+                        GPTModelIntermediateStage,
+                        GPTModelLastStage)
 from .language_model import get_language_model
 from .module import FP16Module
 from .realm_model import ICTBertModel

megatron/model/gpt2_model.py → megatron/model/gpt_model.py

@@ -27,7 +27,7 @@ from .utils import init_method_normal
 from .utils import scaled_init_method_normal
 
 
-def gpt2_attention_mask_func(attention_scores, ltor_mask):
+def gpt_attention_mask_func(attention_scores, ltor_mask):
     attention_scores.masked_fill_(ltor_mask, -10000.0)
     return attention_scores
 
@@ -61,18 +61,18 @@ def post_language_model_processing(lm_output, labels, logit_weights,
         return loss
 
 
-class GPT2ModelBase(MegatronModule):
+class GPTModelBase(MegatronModule):
     """GPT-2 Language model."""
 
     def __init__(self, num_tokentypes=0, parallel_output=True):
-        super(GPT2ModelBase, self).__init__()
+        super(GPTModelBase, self).__init__()
         args = get_args()
 
         self.parallel_output = parallel_output
         self.fp16_lm_cross_entropy = args.fp16_lm_cross_entropy
 
         self.language_model, self._language_model_key = get_language_model(
-            attention_mask_func=gpt2_attention_mask_func,
+            attention_mask_func=gpt_attention_mask_func,
             num_tokentypes=num_tokentypes,
             add_pooler=False,
             init_method=init_method_normal(args.init_method_std),
@@ -81,17 +81,17 @@ class GPT2ModelBase(MegatronModule):
 
         self.initialize_word_embeddings(init_method_normal)
 
-    def forward(self, gpt2_model_input, attention_mask, labels=None,
+    def forward(self, gpt_model_input, 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) = gpt2_model_input
+            (input_ids, position_ids) = gpt_model_input
             args = [input_ids, position_ids, attention_mask]
             kwargs['tokentype_ids'] = tokentype_ids
         else:
-            args = [gpt2_model_input, attention_mask]
+            args = [gpt_model_input, attention_mask]
         lm_output = self.language_model(*args, **kwargs)
 
         if mpu.is_pipeline_last_stage():
@@ -130,17 +130,17 @@ class GPT2ModelBase(MegatronModule):
         self.language_model.load_state_dict(state_dict, strict=strict)
 
 
-class GPT2Model(GPT2ModelBase):
+class GPTModel(GPTModelBase):
 
     def __init__(self, num_tokentypes=0, parallel_output=True):
-        super(GPT2Model, self).__init__(
+        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(GPT2Model, self).forward(
+        return super(GPTModel, self).forward(
             (input_ids, position_ids),
             attention_mask,
             labels=labels,
@@ -150,15 +150,15 @@ class GPT2Model(GPT2ModelBase):
             forward_method_parallel_output=forward_method_parallel_output)
 
 
-class GPT2ModelFirstStage(GPT2ModelBase):
+class GPTModelFirstStage(GPTModelBase):
 
     def __init__(self, num_tokentypes=0):
-        super(GPT2ModelFirstStage, self).__init__(
+        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(GPT2ModelFirstStage, self).forward(
+        return super(GPTModelFirstStage, self).forward(
             (input_ids, position_ids),
             attention_mask,
             tokentype_ids=tokentype_ids,
@@ -166,32 +166,32 @@ class GPT2ModelFirstStage(GPT2ModelBase):
             get_key_value=get_key_value)
 
 
-class GPT2ModelIntermediateStage(GPT2ModelBase):
+class GPTModelIntermediateStage(GPTModelBase):
 
     def __init__(self, num_tokentypes=0):
-        super(GPT2ModelIntermediateStage, self).__init__(
+        super(GPTModelIntermediateStage, self).__init__(
             num_tokentypes=num_tokentypes)
 
     def forward(self, hidden_state, attention_mask,
                 layer_past=None, get_key_value=False):
-        return super(GPT2ModelIntermediateStage, self).forward(
+        return super(GPTModelIntermediateStage, self).forward(
             hidden_state,
             attention_mask,
             layer_past=layer_past,
             get_key_value=get_key_value)
 
 
-class GPT2ModelLastStage(GPT2ModelBase):
+class GPTModelLastStage(GPTModelBase):
 
     def __init__(self, num_tokentypes=0, parallel_output=True):
-        super(GPT2ModelLastStage, self).__init__(
+        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(GPT2ModelLastStage, self).forward(
+        return super(GPTModelLastStage, self).forward(
             hidden_state,
             attention_mask,
             labels=labels,

pretrain_gpt2.py → pretrain_gpt.py

@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Pretrain GPT2"""
+"""Pretrain GPT"""
 
 import torch
 
@@ -22,8 +22,11 @@ from megatron import print_rank_0
 from megatron import get_timers
 from megatron import get_tokenizer
 from megatron import mpu
-from megatron.data.gpt2_dataset import build_train_valid_test_datasets
-from megatron.model import GPT2Model, GPT2ModelFirstStage, GPT2ModelIntermediateStage, GPT2ModelLastStage
+from megatron.data.gpt_dataset import build_train_valid_test_datasets
+from megatron.model import (GPTModel,
+                            GPTModelFirstStage,
+                            GPTModelIntermediateStage,
+                            GPTModelLastStage)
 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
@@ -31,20 +34,20 @@ from megatron.utils import average_losses_across_data_parallel_group
 def model_provider():
     """Build the model."""
 
-    print_rank_0('building GPT2 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 = GPT2ModelFirstStage(num_tokentypes=0)
+            model = GPTModelFirstStage(num_tokentypes=0)
         elif mpu.is_pipeline_last_stage():
-            model = GPT2ModelLastStage(
+            model = GPTModelLastStage(
                 num_tokentypes=0, parallel_output=True)
         else:
-            model = GPT2ModelIntermediateStage(
+            model = GPTModelIntermediateStage(
                 num_tokentypes=0)
     else:
-        model = GPT2Model(num_tokentypes=0, parallel_output=True)
+        model = GPTModel(num_tokentypes=0, parallel_output=True)
 
     return model
 
@@ -124,7 +127,7 @@ def train_valid_test_datasets_provider(train_val_test_num_samples):
     args = get_args()
 
     print_rank_0('> building train, validation, and test datasets '
-                 'for GPT2 ...')
+                 'for GPT ...')
     train_ds, valid_ds, test_ds = build_train_valid_test_datasets(
         data_prefix=args.data_path,
         data_impl=args.data_impl,
@@ -133,7 +136,7 @@ def train_valid_test_datasets_provider(train_val_test_num_samples):
         seq_length=args.seq_length,
         seed=args.seed,
         skip_warmup=(not args.mmap_warmup))
-    print_rank_0("> finished creating GPT2 datasets ...")
+    print_rank_0("> finished creating GPT datasets ...")
 
     return train_ds, valid_ds, test_ds