[fix] Fix unexpected memory cost of gtp-1shot-RS
Open
nemotron
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 6h
- Merged PRs (30d)
- 271
Description
at 1024-scale (TP2), flip gtp1ShotRS from True to False, total device reduced by ~15GB, and max reserved reduced by 3GB.
Goal: figure out why avoid the unexpected memory cost from `total device`
Related [thread](https://nvidia.slack.com/archives/C8HGQL6D8/p1786384942601809) (internal)
Contributor guide
Research direction
Reproduce the 1024-scale (TP2) comparison with gtp1ShotRS set to True and False, measuring total device memory and maximum reserved memory. Trace the code path controlled by this flag to determine why enabling it adds memory, then verify that the unexpected total-device cost is removed without changing the intended behavior.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 42/100