RoPE length extrapolation with interpolation (#7005)

* Push changes

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* Fixes

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* add continue training script

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* [WIP] nonlinear interp

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* Fix

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* override encoder_seq_len

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* Remove nonlinear

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

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

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

* sft with pi (#7006)

* sft with pi

Signed-off-by: Evelina <ebakhturina@nvidia.com>

* update values only if not None"

Signed-off-by: Evelina <ebakhturina@nvidia.com>

---------

Signed-off-by: Evelina <ebakhturina@nvidia.com>

* Address comments

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

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

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

* Add info

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

* Empty

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>

---------

Signed-off-by: MaximumEntropy <sandeep.subramanian.1@umontreal.ca>
Signed-off-by: Evelina <ebakhturina@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Evelina <10428420+ekmb@users.noreply.github.com>
Signed-off-by: Gerald Shen <geshen@nvidia.com>

examples/nlp/language_modeling/conf/megatron_gpt_config.yaml

@@ -83,6 +83,7 @@ model:
   share_embeddings_and_output_weights: True # Share embedding and output layer weights.
   overlap_p2p_comm: False # Overlap p2p communication with computes. This argument is valid only when `virtual_pipeline_model_parallel_size` is larger than 1
   batch_p2p_comm: True # Batch consecutive inter-peer send/recv operations. This argument is valid only when `virtual_pipeline_model_parallel_size` is larger than 1
+  seq_len_interpolation_factor: null # RoPE Interpolation factor for sequence length. This is used to build long-context models with RoPE ex: https://arxiv.org/abs/2306.15595.
 
   tokenizer:
     library: 'megatron'

examples/nlp/language_modeling/megatron_gpt_continue_training.py

@@ -0,0 +1,193 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import tempfile
+
+from omegaconf.omegaconf import OmegaConf, open_dict
+from pytorch_lightning import Trainer
+from pytorch_lightning.plugins.environments import TorchElasticEnvironment
+from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
+
+from nemo.collections.nlp.models.language_modeling.megatron_gpt_model import MegatronGPTModel
+from nemo.collections.nlp.modules.common.megatron.megatron_init import fake_initialize_model_parallel
+from nemo.collections.nlp.parts.nlp_overrides import (
+    GradScaler,
+    MegatronHalfPrecisionPlugin,
+    NLPDDPStrategy,
+    NLPSaveRestoreConnector,
+    PipelineMixedPrecisionPlugin,
+)
+from nemo.core.config import hydra_runner
+from nemo.utils import AppState, logging
+from nemo.utils.exp_manager import exp_manager
+from nemo.utils.model_utils import inject_model_parallel_rank
+
+
+def _modify_config(gpt_cfg, cfg, add_cfg_to_tree=False):
+    """
+    This function modifies the original gpt pre-training config (t5_cfg) with attributes from the finetuning config (cfg).
+    The `add_cfg_to_tree` arg adds `cfg` to the top of the yaml tree which is needed for all `hparams.yaml` files when passed as an arg to `load_from_checkpoint()`.
+    """
+    OmegaConf.set_struct(gpt_cfg, True)
+    OmegaConf.resolve(cfg)
+    with open_dict(gpt_cfg):
+        gpt_cfg.megatron_amp_O2 = cfg.model.get('megatron_amp_O2', False)
+        gpt_cfg.micro_batch_size = cfg.model.micro_batch_size
+        gpt_cfg.global_batch_size = cfg.model.global_batch_size
+        gpt_cfg.sequence_parallel = cfg.model.get("sequence_parallel", False)
+        gpt_cfg.activations_checkpoint_granularity = cfg.model.get("activations_checkpoint_granularity", None)
+        gpt_cfg.activations_checkpoint_num_layers = cfg.model.get("activations_checkpoint_num_layers", None)
+        gpt_cfg.activations_checkpoint_method = cfg.model.get("activations_checkpoint_method", None)
+        gpt_cfg.data = cfg.model.data
+        gpt_cfg.optim = cfg.model.optim
+        gpt_cfg.precision = cfg.trainer.precision
+        gpt_cfg.restore_from_path = cfg.restore_from_path
+        gpt_cfg.resume_from_checkpoint = cfg.model.resume_from_checkpoint
+        gpt_cfg.gradient_as_bucket_view = cfg.model.gradient_as_bucket_view
+        gpt_cfg.encoder_seq_length = cfg.model.encoder_seq_length
+        gpt_cfg.max_position_embeddings = cfg.model.max_position_embeddings
+        gpt_cfg.seq_len_interpolation_factor = cfg.model.seq_len_interpolation_factor
+        gpt_cfg.use_flash_attention = cfg.model.use_flash_attention
+
+        # This is needed when modifying a hparam file directly to load `.ckpt` files.
+        # This is not needed to modify the cfg in `.nemo` files.
+        if add_cfg_to_tree:
+            OmegaConf.resolve(gpt_cfg)
+            gpt_cfg.cfg = gpt_cfg
+
+    return gpt_cfg
+
+
+def load_from_nemo(cls, cfg, trainer, gpt_cfg, modify_confg_fn):
+    gpt_cfg = modify_confg_fn(gpt_cfg, cfg, add_cfg_to_tree=False)
+    save_restore_connector = NLPSaveRestoreConnector()
+    if os.path.isdir(cfg.restore_from_path):
+        save_restore_connector.model_extracted_dir = cfg.restore_from_path
+    model = cls.restore_from(
+        restore_path=cfg.restore_from_path,
+        trainer=trainer,
+        override_config_path=gpt_cfg,
+        save_restore_connector=save_restore_connector,
+    )
+    return model
+
+
+def load_from_checkpoint_dir(cls, cfg, trainer, modify_confg_fn):
+    app_state = AppState()
+    if cfg.model.tensor_model_parallel_size > 1 or cfg.model.pipeline_model_parallel_size > 1:
+        app_state.model_parallel_size = cfg.model.tensor_model_parallel_size * cfg.model.pipeline_model_parallel_size
+        app_state.tensor_model_parallel_size = cfg.model.tensor_model_parallel_size
+        app_state.pipeline_model_parallel_size = cfg.model.pipeline_model_parallel_size
+        (
+            app_state.tensor_model_parallel_rank,
+            app_state.pipeline_model_parallel_rank,
+            app_state.model_parallel_size,
+            app_state.data_parallel_size,
+            app_state.pipeline_model_parallel_split_rank,
+            app_state.virtual_pipeline_model_parallel_rank,
+        ) = fake_initialize_model_parallel(
+            world_size=app_state.model_parallel_size,
+            rank=trainer.global_rank,
+            tensor_model_parallel_size_=cfg.model.tensor_model_parallel_size,
+            pipeline_model_parallel_size_=cfg.model.pipeline_model_parallel_size,
+            pipeline_model_parallel_split_rank_=cfg.model.pipeline_model_parallel_split_rank,
+        )
+    checkpoint_path = inject_model_parallel_rank(
+        os.path.join(cfg.model.pretrained_checkpoint.checkpoint_dir, cfg.model.pretrained_checkpoint.checkpoint_name)
+    )
+    hparams_file = OmegaConf.load(cfg.model.pretrained_checkpoint.hparams_file)
+    gpt_cfg = modify_confg_fn(hparams_file.cfg, cfg, add_cfg_to_tree=True)
+    with tempfile.NamedTemporaryFile(suffix='.yaml') as f:
+        OmegaConf.save(config=gpt_cfg, f=f.name)
+        model = cls.load_from_checkpoint(checkpoint_path=checkpoint_path, trainer=trainer, hparams_file=f.name,)
+        return model
+
+
+def validate_checkpoint_loading_args(cfg):
+    if cfg.checkpoint_dir is None or not os.path.isdir(cfg.checkpoint_dir):
+        raise ValueError(f'Checkpoint directory {cfg.checkpoint_dir} does not exist or is not a directory.')
+    if cfg.checkpoint_name is None:
+        raise ValueError(f'Checkpoint name {cfg.checkpoint_name} is not valid.')
+    if cfg.hparams_file is None or not os.path.isfile(cfg.hparams_file):
+        raise ValueError(f'Hparams file {cfg.hparams_file} does not exist or is not a file.')
+
+
+@hydra_runner(config_path="conf", config_name="megatron_gpt_config")
+def main(cfg) -> None:
+    logging.info("\n\n************** Experiment configuration ***********")
+    logging.info(f'\n{OmegaConf.to_yaml(cfg)}')
+
+    megatron_amp_o2 = cfg.model.get('megatron_amp_O2', False)
+    with_distributed_adam = cfg.model.optim.get('name', 'fused_adam') == 'distributed_fused_adam'
+    plugins = []
+    strategy = NLPDDPStrategy(
+        no_ddp_communication_hook=True,
+        gradient_as_bucket_view=cfg.model.gradient_as_bucket_view,
+        find_unused_parameters=False,
+    )
+    if cfg.trainer.precision in [16, 'bf16']:
+        scaler = None
+        if cfg.trainer.precision == 16:
+            scaler = GradScaler(
+                init_scale=cfg.model.get('native_amp_init_scale', 2 ** 32),
+                growth_interval=cfg.model.get('native_amp_growth_interval', 1000),
+                hysteresis=cfg.model.get('hysteresis', 2),
+            )
+        if megatron_amp_o2 and not with_distributed_adam:
+            plugins.append(MegatronHalfPrecisionPlugin(precision=cfg.trainer.precision, device='cuda', scaler=scaler))
+        else:
+            plugins.append(PipelineMixedPrecisionPlugin(precision=cfg.trainer.precision, device='cuda', scaler=scaler))
+
+    if cfg.get('cluster_type', None) == 'BCP':
+        plugins.append(TorchElasticEnvironment())
+
+    trainer = Trainer(plugins=plugins, strategy=strategy, **cfg.trainer)
+
+    exp_manager(trainer, cfg.exp_manager)
+
+    # update resume from checkpoint found by exp_manager
+    if cfg.model.resume_from_checkpoint is not None:
+        resume_from_checkpoint = cfg.model.resume_from_checkpoint
+    else:
+        resume_from_checkpoint = trainer._checkpoint_connector.resume_from_checkpoint_fit_path
+    logging.info(f'Resuming training from checkpoint: {resume_from_checkpoint}')
+
+    trainer._checkpoint_connector = CheckpointConnector(trainer, resume_from_checkpoint=resume_from_checkpoint)
+
+    if cfg.restore_from_path:
+        save_restore_connector = NLPSaveRestoreConnector()
+        if os.path.isdir(cfg.restore_from_path):
+            save_restore_connector.model_extracted_dir = cfg.restore_from_path
+        gpt_cfg = MegatronGPTModel.restore_from(
+            restore_path=cfg.restore_from_path,
+            trainer=trainer,
+            return_config=True,
+            save_restore_connector=save_restore_connector,
+        )
+        model = load_from_nemo(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)
+    elif cfg.model.get("pretrained_checkpoint", None) is not None:
+        validate_checkpoint_loading_args(cfg.model.pretrained_checkpoint)
+        model = load_from_checkpoint_dir(MegatronGPTModel, cfg, trainer, gpt_cfg, modify_confg_fn=_modify_config)
+    else:
+        print(' > WARNING: No checkpoint provided. Starting from scratch.')
+        # hydra interpolation does not work here as the interpolation key is lost when PTL saves hparams
+        with open_dict(cfg):
+            cfg.model.precision = cfg.trainer.precision
+        model = MegatronGPTModel(cfg.model, trainer)
+    trainer.fit(model)
+
+
+if __name__ == '__main__':
+    main()

examples/nlp/language_modeling/tuning/conf/megatron_gpt_sft.yaml

@@ -60,6 +60,8 @@ model:
   activations_checkpoint_num_layers: null # not used with 'selective'
   answer_only_loss: False # not used right now
   gradient_as_bucket_view: False
+  seq_len_interpolation_factor: null # if not None, seq_len_interpolation_factor will match the base model's value
+  use_flash_attention: null # if not None, will match the base model's value
 
   hidden_dropout: 0.0
   attention_dropout: 0.0

examples/nlp/language_modeling/tuning/megatron_gpt_peft_eval.py

@@ -127,6 +127,10 @@ def main(cfg) -> None:
         peft_model_cfg.data.test_ds = cfg.model.data.test_ds
         peft_model_cfg.activations_checkpoint_granularity = None
         peft_model_cfg.activations_checkpoint_method = None
+        if peft_model_cfg.get("use_flash_attention", False):
+            peft_model_cfg.use_flash_attention = cfg.model.use_flash_attention
+        if cfg.model.get("seq_len_interpolation_factor", None) is not None:
+            peft_model_cfg["seq_len_interpolation_factor"] = cfg.model.seq_len_interpolation_factor
 
     with open_dict(cfg):
         # update the config with the trained model config

examples/nlp/language_modeling/tuning/megatron_gpt_sft.py

@@ -64,6 +64,15 @@ def _modify_config(gpt_cfg, cfg, add_cfg_to_tree=False):
         sft_cls = MegatronGPTSFTModel
         gpt_cfg.target = f"{sft_cls.__module__}.{sft_cls.__name__}"
 
+        if cfg.model.get('use_flash_attention', None) is not None:
+            gpt_cfg.use_flash_attention = cfg.model.use_flash_attention
+
+        if cfg.model.get('seq_len_interpolation_factor', None) is not None:
+            gpt_cfg.seq_len_interpolation_factor = cfg.model.seq_len_interpolation_factor
+
+        sft_cls = MegatronGPTSFTModel
+        gpt_cfg.target = f"{sft_cls.__module__}.{sft_cls.__name__}"
+
         # This is needed when modifying a hparam file directly to load `.ckpt` files.
         # This is not needed to modify the cfg in `.nemo` files.
         if add_cfg_to_tree:

nemo/collections/nlp/models/language_modeling/megatron/gpt_model.py

@@ -166,6 +166,7 @@ def __init__(
         use_emha=False,
         ub_tp_comm_overlap=False,
         use_flash_attention=False,
+        seq_len_interpolation_factor=None,
     ):
         super(GPTModel, self).__init__(share_token_embeddings=share_embeddings_and_output_weights)
 
@@ -249,6 +250,7 @@ def __init__(
             use_emha=use_emha,
             ub_tp_comm_overlap=ub_tp_comm_overlap,
             use_flash_attention=use_flash_attention,
+            seq_len_interpolation_factor=seq_len_interpolation_factor,
         )
 
         if self.share_embeddings_and_output_weights:

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

@@ -249,10 +249,20 @@ def __init__(self, cfg: DictConfig, trainer: Trainer):
             if isinstance(self.model, list):
                 converted_model = []
                 for module in self.model:
-                    converted_model.append(Float16Module(module=module, precision=cfg.precision))
+                    converted_model.append(
+                        Float16Module(
+                            module=module,
+                            precision=cfg.precision,
+                            share_token_embeddings=self.cfg.get('share_embeddings_and_output_weights', True),
+                        )
+                    )
                 self.model = converted_model
             else:
-                self.model = Float16Module(module=self.model, precision=cfg.precision)
+                self.model = Float16Module(
+                    module=self.model,
+                    precision=cfg.precision,
+                    share_token_embeddings=self.cfg.get('share_embeddings_and_output_weights', True),
+                )
 
         if self.trainer.precision == 'bf16':
             self.autocast_dtype = torch.bfloat16
@@ -360,6 +370,7 @@ def model_provider_func(self, pre_process, post_process):
             ub_tp_comm_overlap=self.cfg.get('ub_tp_comm_overlap', False),
             use_flash_attention=self.cfg.get('use_flash_attention', False),
             megatron_legacy=self.cfg.get('megatron_legacy', False),
+            seq_len_interpolation_factor=self.cfg.get('seq_len_interpolation_factor', None),
         )
 
         return model
@@ -981,7 +992,7 @@ def build_pretraining_data_loader(
                     data_parallel_size=parallel_state.get_data_parallel_world_size(),
                     drop_last=drop_last,
                     global_batch_size=self.cfg.global_batch_size,
-                    rampup_batch_size=self.cfg.rampup_batch_size,
+                    rampup_batch_size=self.cfg.get('rampup_batch_size', None),
                     pad_samples_to_global_batch_size=pad_samples_to_global_batch_size,
                 )
             elif self.cfg.data.dataloader_type == 'cyclic':

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

@@ -123,6 +123,7 @@ def get_language_model(
     use_emha=False,
     ub_tp_comm_overlap=False,
     use_flash_attention=False,
+    seq_len_interpolation_factor=None,
 ):
     """Build language model and return along with the key to save."""
 
@@ -200,6 +201,7 @@ def get_language_model(
         use_emha=use_emha,
         ub_tp_comm_overlap=ub_tp_comm_overlap,
         use_flash_attention=use_flash_attention,
+        seq_len_interpolation_factor=seq_len_interpolation_factor,
     )
     # key used for checkpoints.
     language_model_key = 'language_model'
@@ -508,6 +510,7 @@ def __init__(
         use_emha=False,
         ub_tp_comm_overlap=False,
         use_flash_attention=False,
+        seq_len_interpolation_factor=None,
     ):
         super(TransformerLanguageModel, self).__init__(share_token_embeddings=share_embeddings_and_output_weights)
 
@@ -559,7 +562,9 @@ def __init__(
             assert 0 < rotary_percentage <= 1
             if rotary_percentage < 1:
                 rotary_dim = int(rotary_dim * rotary_percentage)
-            self.rotary_pos_emb = RotaryEmbedding(rotary_dim)
+            self.rotary_pos_emb = RotaryEmbedding(
+                rotary_dim, seq_len_interpolation_factor=seq_len_interpolation_factor
+            )
 
         elif position_embedding_type == 'alibi':
             # TODO: If this is used for encoder-decodemax_position_embeddingsr model, implement proper logic and following

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

@@ -254,12 +254,12 @@ def float_conversion(val):
 
 
 class Float16Module(MegatronModule):
-    def __init__(self, module, precision):
+    def __init__(self, module, precision, share_token_embeddings=True):
         if not HAVE_MEGATRON_CORE:
             raise ImportError(
                 "Megatron-core was not found. Please see the NeMo README for installation instructions: https://github.com/NVIDIA/NeMo#megatron-gpt."
             )
-        super().__init__()
+        super().__init__(share_token_embeddings=share_token_embeddings)
         self.precision = precision
 
         if precision == 'bf16':

nemo/collections/nlp/modules/common/megatron/position_embedding/rotary_position_embedding.py

@@ -21,13 +21,28 @@
 
 
 class RotaryEmbedding(nn.Module):
-    def __init__(self, dim):
+    """
+    Implements Rotary Position Embedding from https://arxiv.org/abs/2104.09864.
+    """
+
+    def __init__(self, dim: int, seq_len_interpolation_factor: int = None):
+        """
+        Args:
+
+            dim (int): rotary embedding dimension
+            seq_len_interpolation_factor (int): if not None, discrete positions will be interpolated
+            by this factor via the trick in https://arxiv.org/abs/2306.15595.
+        """
         super().__init__()
+        self.seq_len_interpolation_factor = seq_len_interpolation_factor
         inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
         self.register_buffer('inv_freq', inv_freq)
 
     def forward(self, max_seq_len, offset=0):
         seq = torch.arange(max_seq_len, device=self.inv_freq.device) + offset
+        if self.seq_len_interpolation_factor is not None:
+            seq = seq.type_as(self.inv_freq)
+            seq *= 1 / self.seq_len_interpolation_factor
         freqs = einsum('i , j -> i j', seq.type_as(self.inv_freq), self.inv_freq)
         # first part even vector components, second part odd vector components,
         #  2 * dim in dimension size