Skip to content

ci: add MODEL_FAMILY and TEST_TYPE to test CONFIG blocks#2459

Closed
kajalj22 wants to merge 1 commit into
mainfrom
kajalj/ci-test-metadata-fields
Closed

ci: add MODEL_FAMILY and TEST_TYPE to test CONFIG blocks#2459
kajalj22 wants to merge 1 commit into
mainfrom
kajalj/ci-test-metadata-fields

Conversation

@kajalj22
Copy link
Copy Markdown
Contributor

Summary

  • Adds MODEL_FAMILY and TEST_TYPE fields to every test script's # ===== BEGIN CONFIG ===== block (117 files)
  • MODEL_FAMILY: canonical model identifier (e.g., qwen2.5, llama3.1, gemma3, deepseek-v3)
  • TEST_TYPE: algorithm/task type (e.g., grpo, sft, dpo, distillation, dapo)

Motivation

nemo-ci is adding a standardized test_metadata.json artifact to every CI job (MR !2240). The model_family and test_type fields currently fall back to the full test name (e.g., llm_sft_qwen3_0_6B_1n8g_megatron_yarn_128k) because there are no structured metadata variables in the CONFIG block. This PR adds them so nemo-ci can emit clean, queryable values.

Non-breaking

  • generate_rl_dynamic.sh in nemo-ci parses the CONFIG block with get_cfg_val() which ignores unknown keys
  • Older nemo-ci versions will simply not read these fields
  • No changes to test behavior — these are metadata-only additions

Test plan

  • Verify all 117 scripts have correct MODEL_FAMILY and TEST_TYPE values
  • Validate with nemo-ci's generate_rl_dynamic.sh (will be wired in follow-up nemo-ci MR)

🤖 Generated with Claude Code

@kajalj22 kajalj22 requested a review from a team as a code owner May 11, 2026 05:36
@copy-pr-bot
Copy link
Copy Markdown

copy-pr-bot Bot commented May 11, 2026

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Add structured metadata fields to every test script's CONFIG block
so that nemo-ci can emit clean model_family and test_type values
in test_metadata.json instead of using the full test name.

- MODEL_FAMILY: canonical model identifier (e.g., qwen2.5, llama3.1)
- TEST_TYPE: algorithm/task type (e.g., grpo, sft, dpo, distillation)

These fields are parsed by nemo-ci's generate_rl_dynamic.sh and
passed as CI variables to each test job. Non-breaking: the generator
ignores unknown config keys, so older nemo-ci versions are unaffected.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Kajal Jain <kajalj@nvidia.com>
@kajalj22 kajalj22 force-pushed the kajalj/ci-test-metadata-fields branch from 2b5faad to 91d53c7 Compare May 11, 2026 22:33
@kajalj22
Copy link
Copy Markdown
Contributor Author

Closing — moved to regex-based derivation in nemo-ci's generate_rl_dynamic.sh instead. No RL repo changes needed.

@kajalj22 kajalj22 closed this May 12, 2026
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