NVIDIA-NeMo / NVIDIA-NeMo/Automodel
Add CP to Inkling
@HuiyingLi is already working on this.
Since Jul 30, 2026.
- Dominant language
- Python
- Stars
- 963
- Forks
- 318
- Avg merge
- 3d 20h
- Merged PRs (30d)
- 141
Description
Goal
Add United Ulysses support to the Inkling model family, then progressively extend CP support to other applicable models, prioritizing commonly used MoE models.
Priority and timeline
Phase 1: Inkling — P0
Target completion: August 10, 2026.
- Add end-to-end Ulysses CP support for Inkling.
- Support the attention variants and distributed training configurations used by Inkling.
- Add representative training recipes and automated coverage.
- Validate a 64K sequence-length training configuration.
- Attempt 128K sequence length and document the maximum validated configuration if 128K is not feasible.
Phase 2: Other models
After Inkling is complete, extend CP support incrementally to other applicable models, prioritizing important MoE models.
This phase does not have the one-week deadline. Each model should be handled in a separate follow-up PR and include the same correctness, convergence, and performance evidence described below.
Validation requirements
Convergence + Performance comparison.
Deliverables
- Inkling CP implementation
- Reproducible training recipes on full Inkling-small model
- Convergence plots and raw metrics
- Performance and memory comparison
Assignee
Assigned to: @ghadiaravi13
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.