Skip to content

Fix LUMI container: dclm-core-22, polymath, datasets>=4 - #130

Open
haideraltahan wants to merge 1 commit into
mainfrom
fix-lumi-container
Open

haideraltahan wants to merge 1 commit into
mainfrom
fix-lumi-container

Conversation

@haideraltahan

Copy link
Copy Markdown
Collaborator

Fixes the remaining LUMI container failures after #115.

  • lm-eval[api,tasks] + word2number: polymath crashed on latex2sympy2_extended / word2number
  • datasets>=4.0 (was <4.0.0): matches eval-base; 3.x can't read caches written by 4.x
  • Patch lm-eval for agieval_lsat_ar and squadv2, which crash at n_shot>0 and take the whole dclm-core-22 batch down. The build fails if a patch no longer applies.
  • umask 022 + chmod on /opt/nltk_data, so NLTK data downloads in rootless (proot) builds
  • word2number added to the eval-base extra (lm-eval 0.4.12 doesn't pull it in)

Tested on LUMI dev-g in container mode (pythia-160m, --limit 8):

  • multilingual-oellm-eu: 401/401, all 73 lighteval runs on GPU
  • dclm-core-22: 22/22

- install lm-eval[api,tasks] + word2number (polymath scorer deps)
- datasets>=4.0 to match eval-base; <4 can't read caches written by 4.x
- patch lm-eval for agieval_lsat_ar and squadv2 at n_shot>0
- umask/chmod so NLTK accepts /opt/nltk_data in rootless (proot) builds
- add word2number to the eval-base extra
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.

1 participant