Fix issues with PL 1.8 (NVIDIA#5353)

* Fix issues with PL 1.8

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Set scripting variable

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Fix missing arg

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Cleanup list

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Fix reference

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Try to fix hanging EMA test

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Missing \

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Add strategy

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* See if setting the chdir fixes the hanging DDP test

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* See if removing the subdir setter fixes the issue

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Remove checks

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Try [0,1] for devices

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Add code back

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Remove space

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Update requirements

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Swap import path

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Update references

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Fix deprecated variables

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Fix missing var

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Fix var

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Revert changes

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

* Address review

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>

Signed-off-by: SeanNaren <snarenthiran@nvidia.com>
Co-authored-by: Oleksii Kuchaiev <okuchaiev@users.noreply.github.com>

examples/nlp/language_modeling/megatron_bart_pretraining.py

@@ -13,11 +13,11 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import ModelSummary
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_bart_model import MegatronBARTModel

examples/nlp/language_modeling/megatron_bert_pretraining.py

@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_bert_model import MegatronBertModel

examples/nlp/language_modeling/megatron_ckpt_to_nemo.py

@@ -29,7 +29,7 @@
 
 import torch
 from apex.transformer import parallel_state
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from pytorch_lightning.trainer.trainer import Trainer
 
 from nemo.collections.nlp.models.language_modeling.megatron_bart_model import MegatronBARTModel

examples/nlp/language_modeling/megatron_gpt_pretraining.py

@@ -13,10 +13,10 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_gpt_model import MegatronGPTModel

examples/nlp/language_modeling/megatron_gpt_prompt_learning.py

@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 
 from nemo.collections.nlp.models.language_modeling.megatron_gpt_prompt_learning_model import (
     MegatronGPTPromptLearningModel,

examples/nlp/language_modeling/megatron_retro_cal_shape.py

@@ -12,9 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.plugins.precision.native_amp import NativeMixedPrecisionPlugin
 
 from nemo.collections.nlp.models.language_modeling.megatron_retrieval_model import MegatronRetrievalModel

examples/nlp/language_modeling/megatron_retro_mutransfer_pretrain.py

@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.plugins.precision.native_amp import NativeMixedPrecisionPlugin
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 

examples/nlp/language_modeling/megatron_retro_pretraining.py

@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.plugins.precision.native_amp import NativeMixedPrecisionPlugin
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 

examples/nlp/language_modeling/megatron_t5_lm_adaptation_finetune.py

@@ -13,11 +13,11 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import ModelSummary
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_t5_model import MegatronT5Model

examples/nlp/language_modeling/megatron_t5_pretraining.py

@@ -13,11 +13,11 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import ModelSummary
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_t5_model import MegatronT5Model

examples/nlp/language_modeling/megatron_t5_prompt_learning.py

@@ -12,10 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 
 from nemo.collections.nlp.models.language_modeling.megatron_t5_prompt_learning_model import (
     MegatronT5PromptLearningModel,

examples/nlp/language_modeling/megatron_t5_seq2seq_eval.py

@@ -12,11 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from megatron_t5_seq2seq_finetune import load_from_checkpoint_dir, load_from_nemo, validate_checkpoint_loading_args
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.plugins.precision.native_amp import NativeMixedPrecisionPlugin
 
 from nemo.collections.nlp.models.language_modeling.megatron_finetune_model import MegatronT5FinetuneModel

examples/nlp/language_modeling/megatron_t5_seq2seq_finetune.py

@@ -15,10 +15,10 @@
 import os
 import tempfile
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_finetune_model import MegatronT5FinetuneModel

examples/nlp/language_modeling/tuning/megatron_gpt_adapter_tuning.py

@@ -13,10 +13,10 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 
 from nemo.collections.nlp.models.language_modeling.megatron_gpt_adapter_model import MegatronGPTAdapterLearningModel
 from nemo.collections.nlp.parts.nlp_overrides import (

examples/nlp/language_modeling/tuning/megatron_gpt_ia3_tuning.py

@@ -13,10 +13,10 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 
 from nemo.collections.nlp.models.language_modeling.megatron_gpt_adapter_model import MegatronGPTInfusedAdapterModel
 from nemo.collections.nlp.parts.nlp_overrides import (

examples/nlp/language_modeling/tuning/megatron_t5_adapter_tuning.py

@@ -13,10 +13,10 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 
 from nemo.collections.nlp.models.language_modeling.megatron_t5_adapter_model import MegatronT5AdapterLearningModel
 from nemo.collections.nlp.parts.nlp_overrides import (

examples/nlp/language_modeling/tuning/megatron_t5_ia3_tuning.py

@@ -13,10 +13,10 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 
 from nemo.collections.nlp.models.language_modeling.megatron_t5_adapter_model import MegatronT5InfusedAdapterModel
 from nemo.collections.nlp.parts.nlp_overrides import (

examples/nlp/machine_translation/megatron_nmt_training.py

@@ -13,11 +13,11 @@
 # limitations under the License.
 
 
+from lightning_lite.plugins.environments import TorchElasticEnvironment
 from omegaconf.omegaconf import OmegaConf, open_dict
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import ModelSummary
 from pytorch_lightning.callbacks.timer import Timer
-from pytorch_lightning.plugins.environments.torchelastic_environment import TorchElasticEnvironment
 from pytorch_lightning.trainer.connectors.checkpoint_connector import CheckpointConnector
 
 from nemo.collections.nlp.models.language_modeling.megatron_bart_model import MegatronBARTModel

nemo/collections/nlp/models/dialogue/dialogue_gpt_classification_model.py

@@ -710,7 +710,7 @@ def prepare_data(self):
         self.data_prepared = True
 
     def setup(self, stage=None):
-        super().setup()
+        super().setup(stage)
         if self.cfg.library == "megatron" and self.prompt_learning and stage == "fit":
             if self.cfg.virtual_prompt_style == VirtualPromptStyle.PROMPT_TUNING:
                 self.language_model.init_new_prompts()

nemo/collections/nlp/models/dialogue/dialogue_gpt_generation_model.py

@@ -216,7 +216,7 @@ def mask_and_reduce_loss(self, loss_mask, output_tensor):
         return loss
 
     def setup(self, stage=None):
-        super().setup()
+        super().setup(stage)
         if self.cfg.library == "megatron" and self.prompt_learning:
             self.language_model.init_new_prompts()
 

nemo/collections/nlp/models/dialogue/sgdqa_model.py

@@ -226,7 +226,7 @@ def eval_step_helper(self, batch: List[torch.Tensor]):
         all_start_char_idx = []
         all_end_char_idx = []
 
-        if self.trainer.devices and self.trainer.world_size > 1:
+        if self.trainer.num_devices and self.trainer.world_size > 1:
             world_size = self.trainer.world_size
             for ind in range(world_size):
                 all_example_id_num.append(torch.empty_like(example_id_num))

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

@@ -302,16 +302,16 @@ def on_train_batch_end(self, outputs, batch, batch_idx: int, unused: Optional[in
                 # If the grad scaler skipped its optimizer step due to infs/nans,
                 # decrement the step of all schedulers.
                 if grad_scaler.optimizer_update_skipped is not None and grad_scaler.optimizer_update_skipped is True:
-                    schedulers = self.trainer.lr_schedulers
+                    scheduler_cfgs = self.trainer.lr_scheduler_configs
 
-                    if not schedulers or not self.trainer.lightning_module.automatic_optimization:
+                    if not scheduler_cfgs or not self.trainer.lightning_module.automatic_optimization:
                         return
 
-                    for scheduler in schedulers:
+                    for scheduler_cfg in scheduler_cfgs:
                         # Decrement the counter by 2, then perform a scheduler.step() to perform a no-up
                         # as well as update the optimizer lr in all param groups
-                        scheduler['scheduler'].last_epoch -= 2
-                        scheduler['scheduler'].step()
+                        scheduler_cfg.scheduler.last_epoch -= 2
+                        scheduler_cfg.scheduler.step()
 
                     # Removing the line below because it messes up train_valid_test_num_samples calculation.
                     # self.trainer.fit_loop.max_steps = self.trainer.fit_loop.max_steps + 1

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

@@ -309,16 +309,16 @@ def on_train_batch_end(self, outputs, batch, batch_idx: int, unused: Optional[in
                 # If the grad scaler skipped its optimizer step due to infs/nans,
                 # decrement the step of all schedulers.
                 if grad_scaler.optimizer_update_skipped is not None and grad_scaler.optimizer_update_skipped is True:
-                    schedulers = self.trainer.lr_schedulers
+                    scheduler_cfgs = self.trainer.lr_scheduler_configs
 
-                    if not schedulers or not self.trainer.lightning_module.automatic_optimization:
+                    if not scheduler_cfgs or not self.trainer.lightning_module.automatic_optimization:
                         return
 
-                    for scheduler in schedulers:
+                    for scheduler_cfg in scheduler_cfgs:
                         # Decrement the counter by 2, then perform a scheduler.step() to perform a no-up
                         # as well as update the optimizer lr in all param groups
-                        scheduler['scheduler'].last_epoch -= 2
-                        scheduler['scheduler'].step()
+                        scheduler_cfg.scheduler.last_epoch -= 2
+                        scheduler_cfg.scheduler.step()
 
                     # Increase the max step count by 1
 

nemo/collections/nlp/parts/nlp_overrides.py

@@ -22,16 +22,16 @@
 
 import pytorch_lightning as pl
 import torch
+from lightning_lite.plugins import ClusterEnvironment
+from lightning_lite.utilities.types import _PATH
 from omegaconf import OmegaConf
 from pytorch_lightning.overrides import LightningDistributedModule
-from pytorch_lightning.plugins.environments.cluster_environment import ClusterEnvironment
 from pytorch_lightning.plugins.io.checkpoint_plugin import CheckpointIO
 from pytorch_lightning.plugins.precision.native_amp import NativeMixedPrecisionPlugin
 from pytorch_lightning.strategies.ddp import DDPStrategy
 from pytorch_lightning.trainer.trainer import Trainer
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.fetching import DataFetcher
-from pytorch_lightning.utilities.types import _PATH
 from torch.distributed.algorithms.ddp_comm_hooks.debugging_hooks import noop_hook
 from torch.nn.parallel import DistributedDataParallel
 

nemo/collections/tts/models/fastpitch.py

@@ -19,7 +19,7 @@
 from hydra.utils import instantiate
 from omegaconf import DictConfig, OmegaConf, open_dict
 from pytorch_lightning import Trainer
-from pytorch_lightning.loggers import LoggerCollection, TensorBoardLogger
+from pytorch_lightning.loggers import TensorBoardLogger
 
 from nemo.collections.common.parts.preprocessing import parsers
 from nemo.collections.tts.helpers.helpers import plot_alignment_to_numpy, plot_spectrogram_to_numpy, process_batch
@@ -219,11 +219,10 @@ def tb_logger(self):
             if self.logger is None and self.logger.experiment is None:
                 return None
             tb_logger = self.logger.experiment
-            if isinstance(self.logger, LoggerCollection):
-                for logger in self.logger:
-                    if isinstance(logger, TensorBoardLogger):
-                        tb_logger = logger.experiment
-                        break
+            for logger in self.trainer.loggers:
+                if isinstance(logger, TensorBoardLogger):
+                    tb_logger = logger.experiment
+                    break
             self._tb_logger = tb_logger
         return self._tb_logger
 

nemo/collections/tts/models/radtts.py

@@ -21,7 +21,7 @@
 from hydra.utils import instantiate
 from omegaconf import DictConfig, OmegaConf
 from pytorch_lightning import Trainer
-from pytorch_lightning.loggers import LoggerCollection, TensorBoardLogger
+from pytorch_lightning.loggers import TensorBoardLogger
 
 from nemo.collections.common.tokenizers.text_to_speech.tts_tokenizers import BaseTokenizer
 from nemo.collections.tts.helpers.helpers import plot_alignment_to_numpy
@@ -388,11 +388,10 @@ def tb_logger(self):
             if self.logger is None and self.logger.experiment is None:
                 return None
             tb_logger = self.logger.experiment
-            if isinstance(self.logger, LoggerCollection):
-                for logger in self.logger:
-                    if isinstance(logger, TensorBoardLogger):
-                        tb_logger = logger.experiment
-                        break
+            for logger in self.trainer.loggers:
+                if isinstance(logger, TensorBoardLogger):
+                    tb_logger = logger.experiment
+                    break
             self._tb_logger = tb_logger
         return self._tb_logger
 

nemo/collections/tts/models/tacotron2.py

@@ -20,7 +20,7 @@
 from hydra.utils import instantiate
 from omegaconf import MISSING, DictConfig, OmegaConf, open_dict
 from omegaconf.errors import ConfigAttributeError
-from pytorch_lightning.loggers import LoggerCollection, TensorBoardLogger, WandbLogger
+from pytorch_lightning.loggers import TensorBoardLogger, WandbLogger
 from torch import nn
 
 from nemo.collections.common.parts.preprocessing import parsers
@@ -284,11 +284,10 @@ def validation_step(self, batch, batch_idx):
     def validation_epoch_end(self, outputs):
         if self.logger is not None and self.logger.experiment is not None:
             logger = self.logger.experiment
-            if isinstance(self.logger, LoggerCollection):
-                for logger in self.logger:
-                    if isinstance(logger, TensorBoardLogger):
-                        logger = logger.experiment
-                        break
+            for logger in self.trainer.loggers:
+                if isinstance(logger, TensorBoardLogger):
+                    logger = logger.experiment
+                    break
             if isinstance(logger, TensorBoardLogger):
                 tacotron2_log_to_tb_func(
                     logger, outputs[0].values(), self.global_step, tag="val", log_images=True, add_audio=False,