Safely import optional python packages (#13936)

* Safely import nvidia-resiliency-ext package

The nvidia-resiliency-ext package is not compatible with Windows/MacOS
currently and cannot be installed directly on those platforms. Given the
number of nested imports in the nemo.collections.llm API, it will try and
automatically import the resliency package even if not used. Safely
importing the package allows all modules to be imported without blocking
on the resiliency package for incompatible platforms. This opens up
greater support on more platforms.

If the nvidia-resiliency-ext package is needed when these lines are called,
it will throw an import error only when directly called.

Signed-Off-By: Robert Clark <[email protected]>

* Safely import modelopt algorithms

The default modelopt model transforms rely on the modelopt package being
installed which is not available on all platforms. This module gets
auto-imported when importing the nemo.collections.llm module and will
fail on platforms that don't support modelopt, such as Windows and MacOS,
even if the package isn't used.

Safely importing the package should allow the toolkit to be runnable on
both Windows and MacOS in a minimal state, which increases its
functionality, especially while paired with NeMo-Run.

Signed-Off-By: Robert Clark <[email protected]>

* Apply isort and black reformatting

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

---------

Signed-off-by: Robert Clark <[email protected]>
Signed-off-by: roclark <[email protected]>

Author: roclark

nemo/collections/llm/modelopt/speculative/model_transform.py

@@ -16,17 +16,19 @@
 
 from nemo.collections.llm import GPTModel
 from nemo.utils import logging
-from nemo.utils.import_utils import safe_import
+from nemo.utils.import_utils import UnavailableError, safe_import
 from nemo.utils.model_utils import unwrap_model
 
 mto, HAVE_MODELOPT = safe_import("modelopt.torch.opt")
 mtsp, _ = safe_import("modelopt.torch.speculative")
 
-
-ALGORITHMS = {
-    "eagle3": mtsp.EAGLE3_DEFAULT_CFG,
-    # more TBD
-}
+try:
+    ALGORITHMS = {
+        "eagle3": mtsp.EAGLE3_DEFAULT_CFG,
+        # more TBD
+    }
+except UnavailableError:
+    ALGORITHMS = {}
 
 
 def apply_speculative_decoding(model: nn.Module, algorithm: str = "eagle3") -> nn.Module:

nemo/lightning/nemo_logger.py

@@ -22,12 +22,14 @@
 import lightning.pytorch as pl
 from lightning.pytorch.callbacks.model_checkpoint import ModelCheckpoint as PTLModelCheckpoint
 from lightning.pytorch.loggers import Logger, TensorBoardLogger, WandbLogger
-from nvidia_resiliency_ext.ptl_resiliency.local_checkpoint_callback import LocalCheckpointCallback
 
 from nemo.lightning.io.mixin import IOMixin
 from nemo.lightning.pytorch.callbacks import ModelCheckpoint
 from nemo.utils import logging
 from nemo.utils.app_state import AppState
+from nemo.utils.import_utils import safe_import
+
+LocalCheckpointCallback, HAVE_RES = safe_import('nvidia_resiliency_ext.ptl_resiliency.local_checkpoint_callback')
 
 
 @dataclass