[NeMo-UX] Make TE and Apex dependencies optional by ashors1 · Pull Request #9550 · NVIDIA-NeMo/NeMo

ashors1 · 2024-06-26T22:28:06Z

What does this PR do ?

This PR provides a fallback codepath that defaults to pure Pytorch/jit when TE and Apex are not installed. Note that this is specifically designed to work with NeMo 2.0, so there are no guarantees when running with NeMo 1.0 models.

Collection: [Note which collection this PR will affect]

Changelog

Add specific line by line info of high level changes in this PR.

Usage

You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

Make sure you read and followed Contributor guidelines
Did you write any new necessary tests?
Did you add or update any necessary documentation?
Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
- Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

New Feature
Bugfix
Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

Related to # (issue)

nemo/lightning/apex_utils.py

+# def check_adlr_autoresume_termination(iteration, model, optimizer, lr_scheduler, save: bool):
+#     """Check for autoresume signal and exit if it is received."""
+#     from apex.ppu.checkpointing import save_checkpoint
+#
+#     autoresume = get_adlr_autoresume()
+#     # Add barrier to ensure consistency.
+#     torch.distributed.barrier()
+#     if autoresume.termination_requested():
+#         if save:
+#             save_checkpoint(iteration, model, optimizer, lr_scheduler)
+#         print_rank_0(">>> autoresume termination request found!")
+#         if torch.distributed.get_rank() == 0:
+#             autoresume.request_resume()
+#         print_rank_0(">>> training terminated. Returning")
+#         sys.exit(0)


nemo/lightning/apex_utils.py

nemo/lightning/data.py

nemo/lightning/apex_utils.py

examples/llm/megatron_gpt_pretraining.py

dimapihtar · 2024-07-04T11:40:03Z

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

    HAVE_MEGATRON_CORE = False
+    from typing import Any
+
+    RetroConfig = Any


do we want to do the same for all of others mcore's classes we are trying to import under "try" block?
also, why don't use ApexGuardDefaults?

do we want to do the same for all of others mcore's classes we are trying to import under "try" block?

So this PR is intended to provide a pure pytorch/jit path for NeMo 2.0, so things are not guaranteed to work with NeMo 1.0 models (I have updated the description to reflect this). I added this change because previously, the import of the Retro model in https://github.com/NVIDIA/NeMo/blob/bdb4e89d9ac33d733f8ea7b21552628dda798825/nemo/collections/nlp/models/language_modeling/__init__.py was breaking the 2.0 codepath, but it looks like that's no longer the case anymore. If it makes things clearer, I go ahead and revert this change.

nemo/lightning/data.py

nemo/lightning/megatron_parallel.py

nemo/lightning/pytorch/plugins/data_sampler.py

nemo/collections/llm/gpt/model/base.py

examples/llm/megatron_gpt_pretraining.py

nemo/lightning/data.py

+            setup_microbatch_calculator,
+        )
+
+        HAVE_APEX = True


Signed-off-by: ashors1 <ashors@nvidia.com>