Move pipeline parallel functionality into core with associated changes.

examples/detxoify_lm/finetune_gpt.py

@@ -17,7 +17,8 @@ from megatron import print_rank_0
 from megatron.core import mpu
 from megatron.data.blendable_dataset import BlendableDataset
 from megatron.data.gpt_dataset import build_train_valid_test_datasets
-from megatron.model import GPTModel, ModelType
+from megatron.model import GPTModel
+from megatron.core.enums import ModelType
 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

megatron/core/enums.py

+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+
+import enum
+
+class ModelType(enum.Enum):
+    encoder_or_decoder = 1
+    encoder_and_decoder = 2

megatron/core/parallel_state.py

@@ -58,12 +58,40 @@ def initialize_model_parallel(
     Initialize model data parallel groups.
 
     Arguments:
-        tensor_model_parallel_size: number of GPUs used for tensor model parallelism.
-        pipeline_model_parallel_size: number of GPUs used for pipeline model parallelism.
-        virtual_pipeline_model_parallel_size: number of virtual stages (interleaved
-                                              pipeline).
-        pipeline_model_parallel_split_rank: for models with both encoder and decoder,
-                                            rank in pipeline with split point.
+        tensor_model_parallel_size (int, default = 1):
+            The number of GPUs to split individual tensors across.
+
+        pipeline_model_parallel_size (int, default = 1):
+            The number of tensor parallel GPU groups to split the
+            Transformer layers across. For example, if
+            tensor_model_parallel_size is 4 and
+            pipeline_model_parallel_size is 2, the model will be split
+            into 2 groups of 4 GPUs.
+
+        virtual_pipeline_model_parallel_size (int, optional):
+            The number of stages that each pipeline group will have,
+            interleaving as necessary. If None, no interleaving is
+            performed. For example, if tensor_model_parallel_size is 1,
+            pipeline_model_parallel_size is 4,
+            virtual_pipeline_model_parallel_size is 2, and there are
+            16 transformer layers in the model, the model will be
+            split into 8 stages with two layers each and each GPU
+            would get 2 stages as such (layer number starting with 1):
+
+            GPU 0: [1, 2] [9, 10]
+            GPU 1: [3, 4] [11, 12]
+            GPU 2: [5, 6] [13, 14]
+            GPU 3: [7, 8] [15, 16]
+
+        pipeline_model_parallel_split_rank (int, optional):
+            For models with both an encoder and decoder, the rank in
+            pipeline to switch between encoder and decoder (i.e. the
+            first rank of the decoder). This allows the user to set
+            the pipeline parallel size of the encoder and decoder
+            independently. For example, if
+            pipeline_model_parallel_size is 8 and
+            pipeline_model_parallel_split_rank is 3, then ranks 0-2
+            will be the encoder and ranks 3-7 will be the decoder.
 
     Let's say we have a total of 16 GPUs denoted by g0 ... g15 and we
     use 2 GPUs to parallelize the model tensor, and 4 GPUs to parallelize
@@ -298,8 +326,8 @@ def set_pipeline_model_parallel_rank(rank):
 
 def set_pipeline_model_parallel_split_rank(rank):
     """Set pipeline model parallel split rank."""
-    global _MPU_PIPELINE_MODEL_PARALLEL_SPLIT_RANK
-    _MPU_PIPELINE_MODEL_PARALLEL_SPLIT_RANK = rank
+    global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+    _PIPELINE_MODEL_PARALLEL_SPLIT_RANK = rank
 
 
 def get_tensor_model_parallel_rank():
@@ -318,6 +346,11 @@ def get_pipeline_model_parallel_rank():
     return torch.distributed.get_rank(group=get_pipeline_model_parallel_group())
 
 
+def get_pipeline_model_parallel_split_rank():
+    """Return pipeline model parallel split rank."""
+    global _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+    return _PIPELINE_MODEL_PARALLEL_SPLIT_RANK
+
 
 def is_pipeline_first_stage(ignore_virtual=False):
     """Return True if in the first pipeline model-parallel stage, False otherwise."""

megatron/core/pipeline_parallel/__init__.py

+from .schedules import get_forward_backward_func

megatron/core/tensor_parallel/layers.py

@@ -13,6 +13,8 @@ import torch.nn.functional as F
 import torch.nn.init as init
 from torch.nn.parameter import Parameter
 
+from torch.cuda.amp import custom_fwd, custom_bwd
+
 from megatron.core.parallel_state import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -214,6 +216,7 @@ class LinearWithGradAccumulationAndAsyncCommunication(torch.autograd.Function):
     """See linear_with_grad_accumulation_and_async_allreduce"""
 
     @staticmethod
+    @custom_fwd
     def forward(ctx, input, weight, bias, gradient_accumulation_fusion,
                 async_grad_allreduce, sequence_parallel):
         ctx.save_for_backward(input, weight)
@@ -243,6 +246,7 @@ class LinearWithGradAccumulationAndAsyncCommunication(torch.autograd.Function):
         return output
 
     @staticmethod
+    @custom_bwd
     def backward(ctx, grad_output):
         input, weight = ctx.saved_tensors
         use_bias = ctx.use_bias
@@ -407,8 +411,8 @@ def linear_with_grad_accumulation_and_async_allreduce(
                     "maximum speedup")
                 linear_with_grad_accumulation_and_async_allreduce.warned = True
 
-    with torch.cuda.amp.autocast(enabled=False):
-        return LinearWithGradAccumulationAndAsyncCommunication.apply(*args)
+    return LinearWithGradAccumulationAndAsyncCommunication.apply(*args)
+
 linear_with_grad_accumulation_and_async_allreduce.warned = False
 
 class ColumnParallelLinear(torch.nn.Module):

megatron/core/utils.py

@@ -20,6 +20,21 @@ def divide(numerator, denominator):
     ensure_divisibility(numerator, denominator)
     return numerator // denominator
 
+def get_attr_wrapped_model(model, attr):
+    """Get an attribute from a wrapped model"""
+    if isinstance(model, list):
+        raise RuntimeError("_get_attr_wrapped_model given a list of models")
+
+    while not hasattr(model, attr):
+        if not hasattr(model, "module"):
+            raise RuntimeError(f"_get_attr_wrapped_model couldn't find attribute {attr}")
+
+        model = model.module
+    return getattr(model, attr)
+
+def get_model_type(model):
+    return get_attr_wrapped_model(model, 'model_type')
+
 
 class GlobalMemoryBuffer:
     """Global buffer to avoid dynamic memory allocations.

megatron/model/__init__.py

@@ -8,4 +8,3 @@ from .gpt_model import GPTModel
 from .t5_model import T5Model
 from .language_model import get_language_model
 from .module import Float16Module
-from .enums import ModelType

megatron/model/enums.py

@@ -2,10 +2,6 @@
 
 import enum
 
-class ModelType(enum.Enum):
-    encoder_or_decoder = 1
-    encoder_and_decoder = 2
-
 class LayerType(enum.Enum):
     encoder = 1
     decoder = 2

megatron/model/retro_transformer.py

@@ -20,7 +20,8 @@ from megatron import get_args, get_retro_args, get_tensorboard_writer
 from megatron.core import parallel_state
 from megatron.core import tensor_parallel
 from megatron.core import utils as core_utils
-from megatron.model.enums import AttnMaskType, ModelType, LayerType, AttnType
+from megatron.core.enums import ModelType
+from megatron.model.enums import AttnMaskType, LayerType, AttnType
 from megatron.model import LayerNorm
 from megatron.model.fused_softmax import FusedScaleMaskSoftmax
 from megatron.model.fused_bias_gelu import bias_gelu_impl

megatron/model/transformer.py

@@ -9,7 +9,8 @@ import torch.nn.functional as F
 from megatron import get_timers, get_args, core, get_num_microbatches
 from .module import MegatronModule
 from megatron.core import mpu, tensor_parallel
-from megatron.model.enums import AttnMaskType, ModelType, LayerType, AttnType
+from megatron.core.enums import ModelType
+from megatron.model.enums import AttnMaskType, LayerType, AttnType
 from megatron.model import LayerNorm
 from megatron.model.fused_softmax import FusedScaleMaskSoftmax
 from megatron.model.fused_bias_gelu import bias_gelu_impl

megatron/training.py

@@ -25,8 +25,8 @@ from megatron import print_rank_last
 from megatron.checkpointing import load_checkpoint
 from megatron.checkpointing import save_checkpoint
 from megatron.model import Float16Module
-from megatron.model import ModelType
 from megatron.model import GPTModel
+from megatron.core.enums import ModelType
 from megatron.optimizer import get_megatron_optimizer
 from megatron.initialize import initialize_megatron
 from megatron.initialize import write_args_to_tensorboard
@@ -37,7 +37,7 @@ from megatron.utils import check_adlr_autoresume_termination
 from megatron.utils import unwrap_model
 from megatron.data.data_samplers import build_pretraining_data_loader
 from megatron.utils import calc_params_l2_norm
-from megatron.schedules import get_forward_backward_func
+from megatron.core.pipeline_parallel import get_forward_backward_func
 from megatron.utils import report_memory
 from megatron.model.vision.knn_monitor import compute_feature_bank
 
@@ -400,6 +400,7 @@ def setup_model_and_optimizer(model_provider_func,
     return model, optimizer, opt_param_scheduler
 
 
+
 def train_step(forward_step_func, data_iterator,
                model, optimizer, opt_param_scheduler):
     """Single training step."""
@@ -418,8 +419,16 @@ def train_step(forward_step_func, data_iterator,
     forward_backward_func = get_forward_backward_func()
     fwd_bwd_timers = timers if args.timing_log_level > 1 else None
     losses_reduced = forward_backward_func(
-        forward_step_func, data_iterator, model,
-        optimizer, fwd_bwd_timers, forward_only=False)
+        forward_step_func=forward_step_func,
+        data_iterator=data_iterator,
+        model=model,
+        num_microbatches=get_num_microbatches(),
+        dtype=args.params_dtype,
+        tensor_shape=(args.seq_length, args.micro_batch_size, args.hidden_size),
+        grad_scaler=optimizer.scale_loss,
+        sequence_parallel=args.sequence_parallel,
+        forward_only=False,
+        timers=fwd_bwd_timers)
     timers('forward-backward').stop()
 
     # Empty unused memory.
@@ -794,8 +803,15 @@ def evaluate(forward_step_func,
 
             forward_backward_func = get_forward_backward_func()
             loss_dicts = forward_backward_func(
-                forward_step_func, data_iterator, model, optimizer=None,
-                timers=None, forward_only=True)
+                forward_step_func=forward_step_func,
+                data_iterator=data_iterator,
+                model=model,
+                num_microbatches=get_num_microbatches(),
+                dtype=args.params_dtype,
+                tensor_shape=(args.seq_length, args.micro_batch_size, args.hidden_size),
+                sequence_parallel=args.sequence_parallel,
+                forward_only=True,
+                timers=None)
 
             # Empty unused memory
             if args.empty_unused_memory_level >= 1:

pretrain_bert.py

@@ -11,8 +11,9 @@ from megatron import get_args
 from megatron import print_rank_0
 from megatron import get_timers
 from megatron.core import tensor_parallel
+from megatron.core.enums import ModelType
 from megatron.data.dataset_utils import build_train_valid_test_datasets
-from megatron.model import BertModel, ModelType
+from megatron.model import BertModel
 from megatron.training import pretrain
 from megatron.utils import average_losses_across_data_parallel_group
 

pretrain_gpt.py

@@ -9,8 +9,9 @@ from megatron import print_rank_0
 from megatron import get_timers
 from megatron import get_tokenizer
 from megatron.core import tensor_parallel
+from megatron.core.enums import ModelType
 from megatron.data.gpt_dataset import build_train_valid_test_datasets
-from megatron.model import GPTModel, ModelType
+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

pretrain_ict.py

@@ -13,9 +13,9 @@ from megatron import get_args
 from megatron import print_rank_0
 from megatron import get_timers
 from megatron.core import mpu
+from megatron.core.enums import ModelType
 from megatron.data.biencoder_dataset_utils import get_ict_batch
 from megatron.data.dataset_utils import build_train_valid_test_datasets
-from megatron.model import ModelType
 from megatron.model.biencoder_model import biencoder_model_provider
 from megatron.training import pretrain
 from megatron.utils import average_losses_across_data_parallel_group

pretrain_retro.py

@@ -10,7 +10,8 @@ from megatron import get_timers
 from megatron import get_tokenizer
 from megatron import print_rank_0
 from megatron.core import mpu, tensor_parallel
-from megatron.model import GPTModel, ModelType
+from megatron.core.enums import ModelType
+from megatron.model import GPTModel
 from megatron.training import pretrain
 from megatron.utils import get_ltor_masks_and_position_ids
 from tools.retro.pretraining.retro_dataset import get_retro_datasets

pretrain_t5.py

@@ -12,8 +12,9 @@ from megatron import (
     print_rank_0
 )
 from megatron.core import tensor_parallel
+from megatron.core.enums import ModelType
 from megatron.data.dataset_utils import build_train_valid_test_datasets
-from megatron.model import T5Model, ModelType
+from megatron.model import T5Model
 from megatron.training import pretrain
 from megatron.utils import average_losses_across_data_parallel_group
 

pretrain_vision_classify.py

@@ -6,8 +6,8 @@ import torch
 import torch.nn.functional as F
 from functools import partial
 from megatron import get_args, get_timers, print_rank_0
+from megatron.core.enums import ModelType
 from megatron.data.vit_dataset import build_train_valid_datasets
-from megatron.model import ModelType
 from megatron.model.vision.classification import VitClassificationModel
 from megatron.model.vision.classification import MitClassificationModel
 from megatron.training import pretrain

pretrain_vision_dino.py

@@ -7,6 +7,7 @@ import numpy as np
 import torch.distributed as dist
 from functools import partial
 from megatron import get_args, get_timers, print_rank_0
+from megatron.core.enums import ModelType
 from megatron.data.vit_dataset import build_train_valid_datasets
 from megatron.model.vision.dino import DINOPretrainModel
 from megatron.model.vision.knn_monitor import knn_predict, get_feature_bank
@@ -15,7 +16,6 @@ from megatron.utils import average_losses_across_data_parallel_group, unwrap_mod
 from torch.nn.parallel.distributed import DistributedDataParallel as torchDDP
 from megatron.model import DistributedDataParallel as LocalDDP
 from megatron.model import Float16Module
-from megatron.model import ModelType
 
 def model_provider(pre_process=True, post_process=True):
     """Build the model."""