context parallelism (NVIDIA#7739)

* make nemo recognize sequence_parallel_size

Signed-off-by: xren <[email protected]>

* add helper functions to set up SP running in TE

Signed-off-by: xren <[email protected]>

* slice seq length for a specific rank

Signed-off-by: Xiaowei Ren <[email protected]>

* fix data_parallel_size calculation

Signed-off-by: Xiaowei Ren <[email protected]>

* minor change

Signed-off-by: Xiaowei Ren <[email protected]>

* add missing argument of self

Signed-off-by: Xiaowei Ren <[email protected]>

* pass sp_global_ranks to TE transformer layer

Signed-off-by: Xiaowei Ren <[email protected]>

* fix nsys setting

Signed-off-by: Xiaowei Ren <[email protected]>

* fix seq_len calculation

Signed-off-by: xren <[email protected]>

* fix attn_mask split across seq-length dim

Signed-off-by: xren <[email protected]>

* code update of input split

Signed-off-by: xren <[email protected]>

* fix loss calculation

Signed-off-by: xren <[email protected]>

* fix loss_mask_sum calculation

Signed-off-by: xren <[email protected]>

* fix losss calculation

Signed-off-by: xren <[email protected]>

* rename sequence_parallelism to context_parallelism

Signed-off-by: xren <[email protected]>

* minor change

Signed-off-by: xren <[email protected]>

* fix loss_mask_sum calculation

Signed-off-by: xren <[email protected]>

* make sure do not call megatron-core parallel_state while cp_size is 1

Signed-off-by: xren <[email protected]>

* slice position embedding for different CP rank

Signed-off-by: xren <[email protected]>

* fix mising property decorator

Signed-off-by: xren <[email protected]>

* typo fix

Signed-off-by: xren <[email protected]>

* fix rpe_bias CP slicing

Signed-off-by: xren <[email protected]>

* code style fix

Signed-off-by: xren <[email protected]>

* fix loss_mask_sum calculation

Signed-off-by: xren <[email protected]>

* do not load attention mask if it's not needed

Signed-off-by: Xiaowei Ren <[email protected]>

* bug fix

Signed-off-by: xren <[email protected]>

* fix ubuf size with CP > 1

Signed-off-by: Xiaowei Ren <[email protected]>

* address naming confusion of mixed dp and cp

Signed-off-by: xren <[email protected]>

* rewrite cp code by assuming with_context_parallel=False

Signed-off-by: xren <[email protected]>

* pop context_parallel from dist opt kwargs

Signed-off-by: xren <[email protected]>

* make sure amax reduction group is aware of context parallelism

Signed-off-by: xren <[email protected]>

* remove use_fp8 from initialize_model_parallel

Signed-off-by: xren <[email protected]>

* make implementaitons of setup_transformer_engine_tp_groups and setup_transformer_engine_cp_running consistent

Signed-off-by: xren <[email protected]>

* cp function renaming

Signed-off-by: xren <[email protected]>

* make loss logging broadcast aware of cp

Signed-off-by: xren <[email protected]>

* fix a typo

Signed-off-by: Xiaowei Ren <[email protected]>

* var name fix

Signed-off-by: Xiaowei Ren <[email protected]>

* import transformer layer specs from MCore

Signed-off-by: Xiaowei Ren <[email protected]>

* upgrade MCore version

Signed-off-by: Xiaowei Ren <[email protected]>

* add add context_parallel into the kwargs of dist opt

Signed-off-by: Xiaowei Ren <[email protected]>

* remove redundant cp check

Signed-off-by: Xiaowei Ren <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* code style fix

Signed-off-by: Xiaowei Ren <[email protected]>

* recover docker file

Signed-off-by: Xiaowei Ren <[email protected]>

* fix seq_length of CP

Signed-off-by: Xiaowei Ren <[email protected]>

* recover seq-length which has been fixed in mcore

Signed-off-by: Xiaowei Ren <[email protected]>

* function name fix

Signed-off-by: Xiaowei Ren <[email protected]>

---------

Signed-off-by: xren <[email protected]>
Signed-off-by: Xiaowei Ren <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

nemo/collections/nlp/data/language_modeling/megatron/dataset_utils.py

@@ -1330,7 +1330,7 @@ def get_samples_mapping(
         )
     torch.distributed.barrier()
     counts = torch.cuda.LongTensor([1])
-    torch.distributed.all_reduce(counts, group=parallel_state.get_data_parallel_group())
+    torch.distributed.all_reduce(counts, group=parallel_state.get_data_parallel_group(with_context_parallel=True))
     torch.distributed.all_reduce(counts, group=parallel_state.get_pipeline_model_parallel_group())
     assert counts[0].item() == (
         torch.distributed.get_world_size()

nemo/collections/nlp/data/language_modeling/megatron/gpt_dataset.py

@@ -313,6 +313,7 @@ def __init__(
         self.indexed_dataset = indexed_dataset
         self.drop_last = drop_last
         self.seq_length = seq_length
+        self.get_attention_mask_from_fusion = cfg.get('get_attention_mask_from_fusion', True)
 
         # Checks
         assert np.min(documents) >= 0
@@ -433,13 +434,21 @@ def __getitem__(self, idx):
             logging.debug('Got negative index. Masking loss from this sample')
             loss_mask = torch.zeros_like(loss_mask)
 
-        return {
-            'tokens': tokens,
-            'labels': labels,
-            'attention_mask': attention_mask,
-            'loss_mask': loss_mask,
-            'position_ids': position_ids,
-        }
+        if self.get_attention_mask_from_fusion:
+            return {
+                'tokens': tokens,
+                'labels': labels,
+                'loss_mask': loss_mask,
+                'position_ids': position_ids,
+            }
+        else:
+            return {
+                'tokens': tokens,
+                'labels': labels,
+                'attention_mask': attention_mask,
+                'loss_mask': loss_mask,
+                'position_ids': position_ids,
+            }
 
 
 class MockGPTDataset(Dataset):
@@ -457,6 +466,7 @@ def __init__(
         self.vocab_size = tokenizer.vocab_size
         self.length = num_samples
         self.seed = seed
+        self.get_attention_mask_from_fusion = cfg.get('get_attention_mask_from_fusion', True)
 
         self.attention_mask = torch.tril(torch.ones((self.seq_length, self.seq_length))).unsqueeze(0)
         self.attention_mask = self.attention_mask < 0.5
@@ -476,13 +486,21 @@ def __getitem__(self, idx):
         tokens = torch.from_numpy(np_gen.integers(self.vocab_size, size=[self.seq_length], dtype=np.int64))
         labels = torch.from_numpy(np_gen.integers(self.vocab_size, size=[self.seq_length], dtype=np.int64))
 
-        return {
-            'tokens': tokens,
-            'labels': labels,
-            'attention_mask': self.attention_mask,
-            'loss_mask': self.loss_mask,
-            'position_ids': self.position_ids,
-        }
+        if self.get_attention_mask_from_fusion:
+            return {
+                'tokens': tokens,
+                'labels': labels,
+                'loss_mask': self.loss_mask,
+                'position_ids': self.position_ids,
+            }
+        else:
+            return {
+                'tokens': tokens,
+                'labels': labels,
+                'attention_mask': self.attention_mask,
+                'loss_mask': self.loss_mask,
+                'position_ids': self.position_ids,
+            }
 
 
 @torch.no_grad()
@@ -674,7 +692,7 @@ def _build_index_mappings(
 
     torch.distributed.barrier()
     counts = torch.cuda.LongTensor([1])
-    torch.distributed.all_reduce(counts, group=parallel_state.get_data_parallel_group())
+    torch.distributed.all_reduce(counts, group=parallel_state.get_data_parallel_group(with_context_parallel=True))
     torch.distributed.all_reduce(counts, group=parallel_state.get_pipeline_model_parallel_group())
     assert counts[0].item() == (
         torch.distributed.get_world_size()

nemo/collections/nlp/models/language_modeling/megatron_base_model.py

@@ -169,6 +169,7 @@ def __init__(self, cfg: DictConfig, trainer: Trainer, no_lm_init=True):
             pipeline_model_parallel_size=cfg.get('pipeline_model_parallel_size', 1),
             virtual_pipeline_model_parallel_size=vp_size,
             pipeline_model_parallel_split_rank=cfg.get('pipeline_model_parallel_split_rank', 0),
+            context_parallel_size=cfg.get('context_parallel_size', 1),
             micro_batch_size=cfg.get('micro_batch_size'),
             global_batch_size=cfg.get('global_batch_size'),
             rampup_batch_size=cfg.get('rampup_batch_size', None),
@@ -231,6 +232,27 @@ def setup_transformer_engine_tp_groups(self):
                     tp_group = parallel_state.get_tensor_model_parallel_group()
                     child.set_tensor_parallel_group(tp_group)
 
+    def setup_transformer_engine_cp_groups(self):
+        """ This should be called after context parallel groups have been initialized
+            and only needs to be called when using Transformer Engine.
+        """
+        cp_stream = torch.cuda.Stream()
+
+        for module in self.get_model_module_list():
+            """Set context parallel running
+               Copied from: https://github.com/NVIDIA/TransformerEngine/blob/main/transformer_engine/pytorch/transformer.py
+            """
+            # Deep iterate but skip self to avoid infinite recursion.
+            for index, child in enumerate(module.modules()):
+                if index == 0:
+                    continue
+                if hasattr(child, "set_context_parallel_group"):
+                    child.set_context_parallel_group(
+                        parallel_state.get_context_parallel_group(),
+                        parallel_state.get_context_parallel_global_ranks(),
+                        cp_stream,
+                    )
+
     def _wrap_model_for_O2(self):
         """ Wraps self.model in a float16 wrapper if the model is using megatron amp O2.
             Args:
@@ -556,8 +578,10 @@ def allreduce_gradients(self):
             bucket = buckets[tp]
             grads = [param.grad.data for param in bucket]
             coalesced = torch._utils._flatten_dense_tensors(grads)
-            coalesced /= parallel_state.get_data_parallel_world_size()
-            torch.distributed.all_reduce(coalesced, group=parallel_state.get_data_parallel_group())
+            coalesced /= parallel_state.get_data_parallel_world_size(with_context_parallel=True)
+            torch.distributed.all_reduce(
+                coalesced, group=parallel_state.get_data_parallel_group(with_context_parallel=True)
+            )
             for buf, synced in zip(grads, torch._utils._unflatten_dense_tensors(coalesced, grads)):
                 buf.copy_(synced)
 
@@ -633,7 +657,6 @@ def setup_optimization(
     ):
         optim_kwargs = {} if optim_kwargs is None else optim_kwargs.copy()
         if self.with_distributed_adam:
-
             # Allocate contiguous buffer to avoid extra copies
             optim_kwargs['contiguous_grad_buffer'] = True
 

nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py

@@ -278,7 +278,8 @@ def __init__(self, cfg: DictConfig, trainer: Trainer):
         # Convert the global-batch-based profile index to micro-batch index
         if hasattr(self, '_nsys_profile_enabled'):
             mp_size = cfg.get('tensor_model_parallel_size', 1) * cfg.get('pipeline_model_parallel_size', 1)
-            data_parallel_world_size = trainer.world_size // mp_size
+            cp_size = cfg.get('context_parallel_size', 1)
+            data_parallel_world_size = trainer.world_size // (mp_size * cp_size)
             grad_accum_steps = cfg.get('global_batch_size') // (cfg.get('micro_batch_size') * data_parallel_world_size)
             self._nsys_profile_start_step *= grad_accum_steps
             self._nsys_profile_end_step *= grad_accum_steps
@@ -553,7 +554,9 @@ def initialize_ub_func(self):
             )
 
         input_shape = [
-            self.cfg.get('encoder_seq_length') * self.cfg.get('micro_batch_size'),
+            self.cfg.get('encoder_seq_length')
+            * self.cfg.get('micro_batch_size')
+            // self.cfg.get('context_parallel_size', 1),
             self.cfg.get('hidden_size'),
         ]
 
@@ -834,6 +837,32 @@ def __next__(self):
         # TODO @tmoon: Use once available in Megatron-LM
         # return DataIteratorList(iters)
 
+    def get_batch_on_this_context_parallel_rank(self, batch):
+        cp_size = self.cfg.get('context_parallel_size', 1)
+        num_valid_tokens_in_ub = None
+        if 'loss_mask' in batch and batch['loss_mask'] is not None:
+            num_valid_tokens_in_ub = batch['loss_mask'].sum()
+
+        if cp_size > 1:
+            cp_rank = parallel_state.get_context_parallel_rank()
+            for key, val in batch.items():
+                if val is not None:
+                    seq_dim = 1 if key != 'attention_mask' else 2
+                    val = val.view(
+                        *val.shape[0:seq_dim],
+                        2 * cp_size,
+                        val.shape[seq_dim] // (2 * cp_size),
+                        *val.shape[(seq_dim + 1) :],
+                    )
+                    index = torch.tensor([cp_rank, (2 * cp_size - cp_rank - 1)], device=val.device)
+                    val = val.index_select(seq_dim, index)
+                    val = val.view(*val.shape[0:seq_dim], -1, *val.shape[(seq_dim + 2) :])
+                    batch[key] = val
+
+        batch['num_valid_tokens_in_ub'] = num_valid_tokens_in_ub
+
+        return batch
+
     def get_forward_output_and_loss_func(self, validation_step=False):
         def fwd_output_and_loss_func(dataloader_iter, model, checkpoint_activations_all_layers=None):
 
@@ -852,15 +881,17 @@ def fwd_output_and_loss_func(dataloader_iter, model, checkpoint_activations_all_
                     required_keys.update(('tokens', 'position_ids'))
                 if parallel_state.is_pipeline_last_stage():
                     required_keys.update(('labels', 'loss_mask'))
-            if self.get_attention_mask_from_fusion:
+            if self.get_attention_mask_from_fusion and 'attention_mask' in required_keys:
                 required_keys.remove('attention_mask')
             batch = {key: val.cuda(non_blocking=True) if key in required_keys else None for key, val in batch.items()}
 
+            batch = self.get_batch_on_this_context_parallel_rank(batch)
+
             # Model forward pass
             forward_args = {
                 'input_ids': batch['tokens'],
                 'position_ids': batch['position_ids'],
-                'attention_mask': batch['attention_mask'],
+                'attention_mask': None if self.get_attention_mask_from_fusion else batch['attention_mask'],
                 'labels': batch['labels'],
                 'loss_mask': batch['loss_mask'],
             }
@@ -885,9 +916,10 @@ def fwd_output_and_loss_func(dataloader_iter, model, checkpoint_activations_all_
 
             def loss_func(output_tensor):
                 # Loss for a micro-batch (ub)
-                loss_for_ub = self.loss_func(batch['loss_mask'], output_tensor)
+                loss_for_ub = self.loss_func(batch['loss_mask'], batch['num_valid_tokens_in_ub'], output_tensor)
+                cp_size = self.cfg.get('context_parallel_size', 1)
                 if validation_step and not self.cfg.data.get('validation_drop_last', True):
-                    num_valid_tokens_in_ub = batch['loss_mask'].sum()
+                    num_valid_tokens_in_ub = batch['num_valid_tokens_in_ub']
                     if loss_for_ub.isnan():
                         assert batch['loss_mask'].count_nonzero() == 0, 'Got NaN loss with non-empty input'
                         loss_sum_for_ub = torch.zeros_like(num_valid_tokens_in_ub)
@@ -904,10 +936,10 @@ def loss_func(output_tensor):
                     torch.distributed.all_reduce(
                         loss_sum_and_ub_size_all_gpu, group=parallel_state.get_data_parallel_group()
                     )
-                    return loss_for_ub, {'loss_sum_and_ub_size': loss_sum_and_ub_size_all_gpu}
+                    return loss_for_ub * cp_size, {'loss_sum_and_ub_size': loss_sum_and_ub_size_all_gpu}
                 else:
                     reduced_loss = average_losses_across_data_parallel_group([loss_for_ub])
-                    return loss_for_ub, {'avg': reduced_loss}
+                    return loss_for_ub * cp_size, {'avg': reduced_loss}
 
             return output_tensor, loss_func
 
@@ -1007,10 +1039,11 @@ def on_validation_epoch_end(self):
         if parallel_state.get_pipeline_model_parallel_world_size() > 1:
             if self.loss_broadcast_src_rank is None:
                 dp_size = parallel_state.get_data_parallel_world_size()
+                cp_size = parallel_state.get_context_parallel_world_size()
                 tp_size = parallel_state.get_tensor_model_parallel_world_size()
                 pp_size = parallel_state.get_pipeline_model_parallel_world_size()
-                rank_in_dp_tp_group = torch.distributed.get_rank() % (dp_size * tp_size)
-                last_pipeline_stage_offset = (tp_size * dp_size) * (pp_size - 1)
+                rank_in_dp_tp_group = torch.distributed.get_rank() % (dp_size * cp_size * tp_size)
+                last_pipeline_stage_offset = (tp_size * cp_size * dp_size) * (pp_size - 1)
                 self.loss_broadcast_src_rank = last_pipeline_stage_offset + rank_in_dp_tp_group
             torch.distributed.broadcast(
                 averaged_loss, self.loss_broadcast_src_rank, group=parallel_state.get_pipeline_model_parallel_group(),
@@ -1029,11 +1062,14 @@ def on_test_epoch_end(self):
         logging.info(f'test_loss: {averaged_loss[0]}')
         self.test_step_outputs.clear()  # free memory
 
-    def loss_func(self, loss_mask, output_tensor):
+    def loss_func(self, loss_mask, num_valid_tokens_in_ub, output_tensor):
         losses = output_tensor.float()
         loss_mask = loss_mask.view(-1).float()
         # TODO: add nemo version here
-        loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()  # sequence level nll
+        loss = torch.sum(losses.view(-1) * loss_mask) / num_valid_tokens_in_ub  # sequence level nll
+        cp_size = self.cfg.get('context_parallel_size', 1)
+        if cp_size > 1:
+            torch.distributed.all_reduce(loss, group=parallel_state.get_context_parallel_group())
         return loss
 
     def build_train_valid_test_datasets(self):
@@ -1185,6 +1221,7 @@ def setup(self, stage=None):
 
         if self.cfg.get('transformer_engine', False) or self.cfg.get('mcore_gpt', False):
             self.setup_transformer_engine_tp_groups()
+            self.setup_transformer_engine_cp_groups()
 
     def setup_training_data(self, cfg):
         if hasattr(self, '_train_ds'):
@@ -1243,6 +1280,7 @@ def dummy():
 
             if self.cfg.get('transformer_engine', False):
                 self.setup_transformer_engine_tp_groups()
+                self.setup_transformer_engine_cp_groups()
 
         # set the default sampling params if it is None.
         # default do greedy sampling

nemo/collections/nlp/modules/common/megatron/build_model.py

@@ -151,7 +151,10 @@ def build_model(
         i = torch.cuda.current_device()
         model = [
             torch.nn.parallel.distributed.DistributedDataParallel(
-                model_module, device_ids=[i], output_device=i, process_group=parallel_state.get_data_parallel_group(),
+                model_module,
+                device_ids=[i],
+                output_device=i,
+                process_group=parallel_state.get_data_parallel_group(with_context_parallel=True),
             )
             for model_module in model
         ]

nemo/collections/nlp/modules/common/megatron/language_model.py

@@ -535,6 +535,7 @@ def __init__(
         self.position_embedding_type = position_embedding_type
         self.share_embeddings_and_output_weights = share_embeddings_and_output_weights
         self.sequence_parallel = config.sequence_parallel
+        self.context_parallel = parallel_state.get_context_parallel_world_size() > 1
         if kv_channels is None:
 
             assert (
@@ -722,6 +723,19 @@ def set_input_tensor(self, input_tensor):
 
         self.encoder.set_input_tensor(input_tensor[0])
 
+    def get_position_embedding_on_this_context_parallel_rank(self, position_embedding, seq_dim):
+        cp_size = parallel_state.get_context_parallel_world_size()
+        cp_rank = parallel_state.get_context_parallel_rank()
+        cp_idx = torch.tensor([cp_rank, (2 * cp_size - cp_rank - 1)], device=position_embedding.device)
+        position_embedding = position_embedding.view(
+            *position_embedding.shape[:seq_dim], 2 * cp_size, -1, *position_embedding.shape[(seq_dim + 1) :]
+        )
+        position_embedding = position_embedding.index_select(seq_dim, cp_idx)
+        position_embedding = position_embedding.view(
+            *position_embedding.shape[:seq_dim], -1, *position_embedding.shape[(seq_dim + 2) :]
+        )
+        return position_embedding
+
     def forward(
         self,
         enc_input_ids,
@@ -775,10 +789,16 @@ def forward(
             else:
                 enc_seq_length = encoder_input.size(0)
 
+        if self.context_parallel:
+            enc_seq_length = enc_seq_length * parallel_state.get_context_parallel_world_size()
+
         rotary_pos_emb = None
         encoder_self_attention_relative_position_bias = None
         if self.position_embedding_type == 'rope':
             rotary_pos_emb = self.rotary_pos_emb(enc_seq_length)
+
+            if self.context_parallel:
+                rotary_pos_emb = self.get_position_embedding_on_this_context_parallel_rank(rotary_pos_emb, 0)
         elif (
             self.position_embedding_type == 'alibi'
             or self.position_embedding_type == 'sandwich'
@@ -790,6 +810,11 @@ def forward(
             # causal attention bias: [1, head, 1, k]
             # non-causal attention bias: [1, head, q, k]
 
+            if self.context_parallel and encoder_self_attention_relative_position_bias.shape[-2] > 1:
+                encoder_self_attention_relative_position_bias = self.get_position_embedding_on_this_context_parallel_rank(
+                    encoder_self_attention_relative_position_bias, 2
+                )
+
         # encoder.
         if enc_hidden_states is None:
             encoder_output = self.encoder(