Skip to content

fix: fix lite csa.py code doesn't match megatron-lm dev rescontruct code - #29

Open
chenjiaoAngel wants to merge 2 commits into
verl-project:litefrom
chenjiaoAngel:lite_fix_match_dev
Open

fix: fix lite csa.py code doesn't match megatron-lm dev rescontruct code#29
chenjiaoAngel wants to merge 2 commits into
verl-project:litefrom
chenjiaoAngel:lite_fix_match_dev

Conversation

@chenjiaoAngel

Copy link
Copy Markdown

No description provided.

@ISEEKYAN

Copy link
Copy Markdown

Reviewed while bundling the open lite PRs for the upstream sync (NVIDIA#6397).

The change itself is correct, and it is a hard prerequisite for the sync — but the landing order matters.

Correct: nv/dev restructured CSA in 1c44a5709 ("Refactor CSA structure: Move CSA implementation helpers into csa_utils dir", NVIDIA#6372). On current dev, csa_cp_utils.py and csa_cp_layout_kernels.py no longer exist; the helpers now live in csa_utils/{cp_utils,cp_layout_kernels}.py. The new imports here match that layout exactly, so lite's csa.py cannot import against current dev without this PR.

Ordering hazard: our validation environment pins mcore at 43124b60c (2026-08-08), which I checked and it is still the old layout (csa_cp_utils.py / csa_cp_layout_kernels.py present, no csa_utils/). Merging this into lite on its own makes csa.py importable only against the newer dev, so every CSA test in the standard profile fails with ImportError against the pinned mcore.

So this should land together with bumping the pinned mcore to a commit that contains 1c44a5709, as one step: bump pin -> merge this -> re-run the standard profile. Otherwise we lose CSA test coverage in the window between.

Also: the branch is based on 5e9a17e4d (2026-08-11), which is 22 commits behind lite. Please rebase before merge so what lands is the current tree.

No objection to the content — just do not merge it in isolation.

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