Skip to content

feat(convert): add distributed multi-GPU checkpoint conversion#238

Open
rapaul-nv wants to merge 3 commits into
mainfrom
rapaul/playbookv3-5
Open

feat(convert): add distributed multi-GPU checkpoint conversion#238
rapaul-nv wants to merge 3 commits into
mainfrom
rapaul/playbookv3-5

Conversation

@rapaul-nv
Copy link
Copy Markdown
Contributor

Run HF↔Megatron conversion through Megatron-Bridge's multi-GPU converter by default so large checkpoints no longer need to materialize on one GPU.

  • add distributed torchrun command construction with tp/pp/ep/etp validation
  • support config-owned converter images/scripts while preserving env profile logistics
  • add Super-compatible tiny conversion configs for the 26.02 stack
  • document distributed conversion knobs and add focused runner/config-loader tests

Run HF↔Megatron conversion through Megatron-Bridge's multi-GPU converter by default so large checkpoints no longer need to materialize on one GPU.

- add distributed torchrun command construction with tp/pp/ep/etp validation
- support config-owned converter images/scripts while preserving env profile logistics
- add Super-compatible tiny conversion configs for the 26.02 stack
- document distributed conversion knobs and add focused runner/config-loader tests

Signed-off-by: Rakesh Paul <rapaul@nvidia.com>
@rapaul-nv rapaul-nv marked this pull request as ready for review May 30, 2026 09:13
@rapaul-nv rapaul-nv requested a review from rkalaniNV May 30, 2026 10:28
Comment thread docs/train-models/reference/convert/megatron-to-hf.md Outdated
Comment thread src/nemotron/steps/convert/megatron_to_hf/README.md
…ation

Eliminate the `script.path` option from the documentation for both HF to Megatron and Megatron to HF conversion steps, as it is no longer necessary. This change simplifies the documentation and aligns it with the current implementation.

Signed-off-by: Rakesh Paul <rapaul@nvidia.com>
@rapaul-nv rapaul-nv requested a review from rkalaniNV May 30, 2026 11:05
- Renamed test functions for better readability and alignment with their purpose.
- Removed unnecessary assertions and parameters to streamline tests.
- Updated expected configuration values to reflect current implementation standards.

Signed-off-by: Rakesh Paul <rapaul@nvidia.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants