Merge branch 'transformer_engine_rebase' into 'main'

Transformer Engine Integration Rebase

See merge request ADLR/megatron-lm!487

megatron/arguments.py

@@ -28,6 +28,7 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False):
     parser = _add_vision_args(parser)
     parser = _add_logging_args(parser)
     parser = _add_inference_args(parser)
+    parser = _add_transformer_engine_args(parser)
 
     # Custom arguments.
     if extra_args_provider is not None:
@@ -304,6 +305,18 @@ def validate_args(args, defaults={}):
             'v1.10 and above (Nvidia Pytorch container >= 21.07). Current ' \
             'pytorch version is v%s.%s.' % (TORCH_MAJOR, TORCH_MINOR)
 
+    # Tranformer-Engine/FP8 related checking
+    if args.fp8_e4m3 or args.fp8_hybrid:
+        assert args.transformer_impl == 'transformer_engine', \
+            'transformer-engine required for fp8 training and inference'
+
+    assert not (args.fp8_e4m3 and args.fp8_hybrid), \
+        'cannot train with both fp8 e4m3 and hybrid formatting'
+
+    if args.fp16:
+        assert args.transformer_impl == 'local', \
+            'transformer-engine not yet approved for fp16 training and inference'
+
     if args.recompute_granularity == 'selective':
         assert args.recompute_method is None, \
             'recompute method is not yet supported for ' \
@@ -355,6 +368,33 @@ def _check_arg_is_not_none(args, arg):
     assert getattr(args, arg) is not None, '{} argument is None'.format(arg)
 
 
+def _add_transformer_engine_args(parser):
+    group = parser.add_argument_group(title='Transformer-Engine')
+
+    group.add_argument('--fp8-e4m3', action='store_true',
+                        help='E4M3 TransformerLayer', dest='fp8_e4m3')
+    group.add_argument('--fp8-hybrid', action='store_true',
+                        help='Hybrid FP8 TransformerLayer', dest='fp8_hybrid')
+    group.add_argument('--no-fp8-wgrad', action='store_false',
+                        help='Execute wgrad in higher precision even for FP8 runs', dest='fp8_wgrad')
+    group.add_argument('--fp8-margin', type=int, default=0,
+                        help='Scaling margin for fp8', dest='fp8_margin')
+    group.add_argument('--fp8-interval', type=int, default=1,
+                        help='Scaling update interval for fp8', dest='fp8_interval')
+    group.add_argument('--transformer-impl', default='local',
+                       choices=['local', 'transformer_engine'],
+                       help='Which Transformer implementation to use.',
+                       dest='transformer_impl')
+    group.add_argument('--fp8-amax-history-len', type=int, default=1,
+                        help='Number of steps for which amax history is recorded per tensor',
+                        dest='fp8_amax_history_len')
+    group.add_argument('--fp8-amax-compute-algo', default='most_recent',
+                       choices=['most_recent', 'max'],
+                       help='Algorithm for computing amax from history',
+                       dest='fp8_amax_compute_algo')
+
+    return parser
+
 def _add_inference_args(parser):
     group = parser.add_argument_group(title='inference')
 

megatron/fused_kernels/__init__.py

@@ -18,11 +18,14 @@ def load(args):
 
     # Check if cuda 11 is installed for compute capability 8.0
     cc_flag = []
-    _, bare_metal_major, _ = _get_cuda_bare_metal_version(
+    _, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(
         cpp_extension.CUDA_HOME)
     if int(bare_metal_major) >= 11:
         cc_flag.append('-gencode')
         cc_flag.append('arch=compute_80,code=sm_80')
+        if int(bare_metal_minor) >= 7:
+            cc_flag.append('-gencode')
+            cc_flag.append('arch=compute_90,code=sm_90')
 
     # Build path
     srcpath = pathlib.Path(__file__).parent.absolute()
@@ -75,11 +78,14 @@ def load(args):
     # Mixed precision fused layer norm.
     # =================================
 
+    extra_hopper_flags = ['-U__CUDA_NO_HALF_OPERATORS__',
+                          '-U__CUDA_NO_HALF_CONVERSIONS__']
+
     extra_cuda_flags = ['-maxrregcount=50']
     sources=[srcpath / 'layer_norm_cuda.cpp',
              srcpath / 'layer_norm_cuda_kernel.cu']
     fused_mix_prec_layer_norm_cuda = _cpp_extention_load_helper(
-        "fused_mix_prec_layer_norm_cuda", sources, extra_cuda_flags)
+        "fused_mix_prec_layer_norm_cuda", sources, extra_cuda_flags + extra_hopper_flags)
 
     # =================================
     # Fused gradient accumulation to weight gradient computation of linear layer
@@ -89,7 +95,7 @@ def load(args):
         sources=[srcpath / 'fused_weight_gradient_dense.cpp',
                  srcpath / 'fused_weight_gradient_dense.cu']
         fused_dense_cuda = _cpp_extention_load_helper(
-            "fused_dense_cuda", sources, [])
+            "fused_dense_cuda", sources, extra_hopper_flags)
 
 
 def _get_cuda_bare_metal_version(cuda_dir):

megatron/model/transformer.py

@@ -6,7 +6,7 @@ from contextlib import nullcontext
 import torch
 import torch.nn.functional as F
 
-from megatron import get_timers, get_args, core
+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
@@ -15,7 +15,6 @@ from megatron.model.fused_softmax import FusedScaleMaskSoftmax
 from megatron.model.fused_bias_gelu import bias_gelu_impl
 from megatron.model.utils import attention_mask_func, openai_gelu, erf_gelu
 
-
 """ We use the following notation throughout this file:
      h: hidden size
      n: number of attention heads
@@ -810,6 +809,7 @@ class ParallelTransformer(MegatronModule):
         self.post_process = post_process
         self.input_tensor = None
         self.drop_path_rate = drop_path_rate
+        self.transformer_impl = args.transformer_impl
 
         # Store activation checkpoiting flag.
         self.recompute_granularity = args.recompute_granularity
@@ -820,6 +820,31 @@ class ParallelTransformer(MegatronModule):
 
         self.sequence_parallel = args.sequence_parallel
 
+        # Transformer Engine Init.
+        if self.transformer_impl == 'transformer_engine':
+            global transformer_engine
+            import transformer_engine
+        self.use_fp8 = args.fp8_e4m3 or args.fp8_hybrid
+        self.fp8_recipe = None
+        self.fp8_group = mpu.get_data_parallel_group()
+        if self.use_fp8:
+            if args.fp8_e4m3:
+                fp8_format = transformer_engine.common.recipe.Format.E4M3
+            elif args.fp8_hybrid:
+                fp8_format = transformer_engine.common.recipe.Format.HYBRID
+            self.fp8_recipe = transformer_engine.common.recipe.DelayedScaling(
+                margin=args.fp8_margin,
+                interval=args.fp8_interval,
+                fp8_format=fp8_format,
+                amax_history_len=args.fp8_amax_history_len,
+                amax_compute_algo=args.fp8_amax_compute_algo,
+                override_linear_precision=(False, False, not args.fp8_wgrad),
+            )
+
+        self.num_microbatches_in_previous_step = -1
+        self.microbatch_count = 0
+        self.checkpoint_core_attention = args.recompute_granularity == 'selective'
+
         # Number of layers.
         self.num_layers = _get_num_layers(
             args,
@@ -830,6 +855,7 @@ class ParallelTransformer(MegatronModule):
 
         # Transformer layers.
         def build_layer(layer_number):
+            if args.transformer_impl == 'local':
                 return ParallelTransformerLayer(
                     init_method,
                     output_layer_init_method,
@@ -837,6 +863,35 @@ class ParallelTransformer(MegatronModule):
                     layer_type=layer_type,
                     self_attn_mask_type=self_attn_mask_type,
                     drop_path_rate=self.drop_path_rates[layer_number - 1])
+            else:
+                return transformer_engine.pytorch.TransformerLayer(
+                    args.hidden_size,
+                    args.ffn_hidden_size,
+                    args.num_attention_heads,
+                    layernorm_epsilon=args.layernorm_epsilon,
+                    hidden_dropout=args.hidden_dropout,
+                    attention_dropout=args.attention_dropout,
+                    init_method=init_method,
+                    output_layer_init_method=output_layer_init_method,
+                    layer_number=layer_number,
+                    kv_channels=args.kv_channels,
+                    self_attn_mask_type=self_attn_mask_type.name,
+                    tp_group=mpu.get_tensor_model_parallel_group(),
+                    get_rng_state_tracker=tensor_parallel.get_cuda_rng_tracker,
+                    fuse_wgrad_accumulation=args.gradient_accumulation_fusion,
+                    apply_query_key_layer_scaling=args.apply_query_key_layer_scaling,
+                    attention_softmax_in_fp32=args.attention_softmax_in_fp32,
+                    seq_length=args.seq_length,
+                    micro_batch_size=args.micro_batch_size,
+                    sequence_parallel=args.sequence_parallel,
+                    params_dtype=args.params_dtype,
+                    apply_residual_connection_post_layernorm=args.apply_residual_connection_post_layernorm,
+                    output_layernorm=False,
+                    layer_type="encoder",
+                    drop_path_rate=self.drop_path_rates[layer_number - 1],
+                    set_parallel_mode=True,
+                    fuse_qkv_params=True)
+
         if args.virtual_pipeline_model_parallel_size is not None:
             assert args.num_layers % args.virtual_pipeline_model_parallel_size == 0, \
                 'num_layers_per_stage must be divisible by ' \
@@ -896,19 +951,20 @@ class ParallelTransformer(MegatronModule):
         return self.layers[layer_number]
 
     def _checkpointed_forward(self, hidden_states, attention_mask,
-                              encoder_output, enc_dec_attn_mask):
+                              encoder_output, enc_dec_attn_mask, is_first_microbatch):
         """Forward method with activation checkpointing."""
-        def custom(start, end):
-            def custom_forward(*inputs):
-                x_ = inputs[0]
-                attention_mask = inputs[1]
-                encoder_output = inputs[2]
-                enc_dec_attn_mask = inputs[3]
+        def custom(start, end, is_transformer_engine=False):
+            def custom_forward(*args, **kwargs):
                 for index in range(start, end):
                     layer = self._get_layer(index)
-                    x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask)
+                    x_ = layer(*args, **kwargs)
                 return x_
+            def custom_forward_transformer_engine(*args, **kwargs):
+                return custom_forward(*args, is_first_microbatch=is_first_microbatch, **kwargs)
+            if not is_transformer_engine:
                 return custom_forward
+            else:
+                return custom_forward_transformer_engine
 
         if self.recompute_method == 'uniform':
             # Uniformly divide the total number of Transformer layers and checkpoint
@@ -916,10 +972,19 @@ class ParallelTransformer(MegatronModule):
             # A method to further reduce memory usage reducing checkpoints.
             l = 0
             while l < self.num_layers:
+                if self.transformer_impl == 'transformer_engine':
+                    hidden_states = transformer_engine.pytorch.distributed.checkpoint(
+                        custom(l, l + self.recompute_num_layers, is_transformer_engine=True),
+                        self.distribute_saved_activations,
+                        tensor_parallel.get_cuda_rng_tracker,
+                        mpu.get_tensor_model_parallel_group(),
+                        hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                else:
                     hidden_states = tensor_parallel.checkpoint(
                         custom(l, l + self.recompute_num_layers),
                         self.distribute_saved_activations,
                         hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+
                 l += self.recompute_num_layers
 
         elif self.recompute_method == 'block':
@@ -928,10 +993,22 @@ class ParallelTransformer(MegatronModule):
             # A method fully use the device memory removing redundant re-computation.
             for l in range(self.num_layers):
                 if l < self.recompute_num_layers:
+                    if self.transformer_impl == 'transformer_engine':
+                        hidden_states = transformer_engine.pytorch.distributed.checkpoint(
+                            custom(l, l + 1, is_transformer_engine=True),
+                            self.distribute_saved_activations,
+                            tensor_parallel.get_cuda_rng_tracker,
+                            mpu.get_tensor_model_parallel_group(),
+                            hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                    else:
                         hidden_states = tensor_parallel.checkpoint(
                             custom(l, l + 1),
                             self.distribute_saved_activations,
                             hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
+                else:
+                    if self.transformer_impl == 'transformer_engine':
+                        hidden_states = custom(l, l + 1, is_transformer_engine=True)(
+                            hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
                     else:
                         hidden_states = custom(l, l + 1)(
                             hidden_states, attention_mask, encoder_output, enc_dec_attn_mask)
@@ -991,21 +1068,48 @@ class ParallelTransformer(MegatronModule):
             rng_context = nullcontext()
 
         with rng_context:
+            # The fp8_autocast context manager is a no-op when enabled=True
+            # The if...else serves to short circuit name resolution for fp8_autocast
+            with transformer_engine.pytorch.fp8_autocast(
+                enabled=self.use_fp8,
+                fp8_recipe=self.fp8_recipe,
+                fp8_group=self.fp8_group
+            ) if self.use_fp8 else nullcontext():
+                # Determine if the current iteration is first microbatch
+                if self.num_microbatches_in_previous_step != get_num_microbatches():
+                    self.microbatch_count = 0 # Reset count on new batch size rampup interval
+                self.num_microbatches_in_previous_step = get_num_microbatches()
+                is_first_microbatch = self.microbatch_count % get_num_microbatches() == 0
+
                 # Forward pass.
                 if self.recompute_granularity == 'full':
                     hidden_states = self._checkpointed_forward(hidden_states,
                                                                attention_mask,
                                                                encoder_output,
-                                                           enc_dec_attn_mask)
+                                                               enc_dec_attn_mask,
+                                                               is_first_microbatch)
                 else:
+                    forward_kwargs = {
+                        'encoder_output': encoder_output,
+                        'enc_dec_attn_mask': enc_dec_attn_mask,
+                        'inference_params': inference_params,
+                    }
+
+                    if self.transformer_impl == 'transformer_engine':
+                        forward_kwargs['is_first_microbatch'] = is_first_microbatch
+                        forward_kwargs['checkpoint_core_attention'] = self.checkpoint_core_attention
+
                     for index in range(self.num_layers):
                         layer = self._get_layer(index)
+
                         hidden_states = layer(
                             hidden_states,
                             attention_mask,
-                        encoder_output=encoder_output,
-                        enc_dec_attn_mask=enc_dec_attn_mask,
-                        inference_params=inference_params)
+                            **forward_kwargs)
+
+                # Skip counter update for eval and activation checkpointing
+                if torch.is_grad_enabled() and self.training:
+                    self.microbatch_count += 1
 
         # Final layer norm.
         if self.post_process and self.post_layer_norm:

megatron/training.py

@@ -26,6 +26,7 @@ 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.optimizer import get_megatron_optimizer
 from megatron.initialize import initialize_megatron
 from megatron.initialize import write_args_to_tensorboard
@@ -251,6 +252,12 @@ def get_model(model_provider_func, model_type=ModelType.encoder_or_decoder, wrap
     if not isinstance(model, list):
         model = [model]
 
+    # Disallow training and inference with Transformer Engine
+    # for non-GPT models
+    args.allow_transformer_engine = all([type(m) == GPTModel for m in model])
+    assert args.allow_transformer_engine or args.transformer_impl == 'local', \
+        'Transformer Engine is only approved for GPT models'
+
     # Set tensor model parallel attributes if not set.
     # Only parameters that are already tensor model parallel have these
     # attributes set for them. We should make sure the default attributes

pretrain_gpt.py

@@ -113,4 +113,6 @@ if __name__ == "__main__":
 
     pretrain(train_valid_test_datasets_provider, model_provider,
              ModelType.encoder_or_decoder,
-             forward_step, args_defaults={'tokenizer_type': 'GPT2BPETokenizer'})
+             forward_step,
+             args_defaults={'tokenizer_type': 'GPT2BPETokenizer'}
+    )