Skip to content

NemotronH: secondary perf/diagnostic patch set (Liger, TP CE, FLOPs, benchmark) - #1

Open
jasont314 wants to merge 3 commits into
pr1-nemotron-pp-ep-squadfrom
pr2-nemotron-secondary-patches
Open

NemotronH: secondary perf/diagnostic patch set (Liger, TP CE, FLOPs, benchmark)#1
jasont314 wants to merge 3 commits into
pr1-nemotron-pp-ep-squadfrom
pr2-nemotron-secondary-patches

Conversation

@jasont314

Copy link
Copy Markdown
Owner

What does this PR do ?

Adds the secondary (auxiliary) NemotronH patch set that is separable from the core PP/EP + SQuAD correctness PR, focused on performance paths, diagnostics, and benchmark/reporting fidelity.

Changelog

  • Add NemotronH-specific Liger patch path in:
    • nemo_automodel/_transformers/auto_model.py
  • Improve NCCL distributed init robustness by binding device_id during PG init in:
    • nemo_automodel/components/distributed/init_utils.py
  • Add optional TP-aware parallel cross-entropy path for DTensor vocab-sharded logits in:
    • nemo_automodel/components/loss/masked_ce.py
  • Add NemotronH FLOPs accounting formula and mapping for improved MFU analysis in:
    • nemo_automodel/components/utils/flops_utils.py
  • Update benchmark runtime behavior in:
    • nemo_automodel/recipes/llm/benchmark.py
    • scoped PP loss sync helper usage
    • TF32 enablement for benchmark throughput on supported GPUs
  • Add secondary patch documentation:
    • NEMOTRON_SECONDARY_PATCH_NOTES.md

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?

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

Additional Information

  • Related to # (issue)
  • This is a stacked follow-up PR on top of pr1-nemotron-pp-ep-squad.
  • After PR1 merges, this PR base should be retargeted to NVIDIA-NeMo/Automodel:main.
  • Patch details and risk notes: NEMOTRON_SECONDARY_PATCH_NOTES.md

ZhiyuLi-Nvidia and others added 3 commits January 22, 2026 09:01
Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

perf lab

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

perf lab (ep)

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

better rope cache

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

arithmetic intensity lab

arithmetic intensity lab fix

arithmetic intensity lab fix

arithmetic intensity lab fix

arithmetic intensity lab fix

Revert "arithmetic intensity lab fix"

This reverts commit b9a2c57.

arithmetic intensity lab fix

arithmetic intensity answer

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

mixtral perf lab

mixtral perf lab

mixtral perf lab fix

mixtral perf lab fix

mixtral perf lab fix

mixtral perf lab fix

update grade_assignment

update lab

update

Signed-off-by: Zhiyu Li <zhiyul@nvidia.com>

arithmetic intensity answer

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>

remove reference

remove reference

Signed-off-by: Zhiyu Li <zhiyul@NVIDIA.com>
Integrates NemotronH PP/EP execution and safety guards, fixes fixed-length SQuAD label/tokenization masking so num_label_tokens stays > 0, and adds PP schedule/device/logging robustness updates. Includes optimized PP+EP SQuAD config, patch notes, training artifacts (baseline/optimized JSONL), and unit-test updates (79 passed, 5 skipped).
…benchmark)

Adds auxiliary optimization and reporting updates separated from core PP/EP/SQuAD correctness path.
@jasont314
jasont314 force-pushed the pr1-nemotron-pp-ep-squad branch 5 times, most recently from 2a6fee4 to bf35003 Compare February 20, 2026 12:39
@jasont314
jasont314 force-pushed the pr1-nemotron-pp-ep-squad branch from d38c94b to 29f3abe Compare February 26, 2026 20:10
@jasont314
jasont314 force-pushed the pr1-nemotron-pp-ep-squad branch 3 times, most recently from d882293 to 8bb02c7 Compare March 17, 2026 07:34
@jasont314
jasont314 force-pushed the pr1-nemotron-pp-ep-squad branch from 8bb02c7 to b39273b Compare March 25, 2026 10:53
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