Skip to content

launch: give RL islands the env Megatron actually wants (re-land #63 on main) - #68

Merged
michaellchung merged 3 commits into
mainfrom
fix/sky-island-tp-env
Sep 29, 2026
Merged

michaellchung merged 3 commits into
mainfrom
fix/sky-island-tp-env

Conversation

@michaellchung

Copy link
Copy Markdown
Contributor

#63 was merged into its stacked base branch after that branch had already landed on main, so its changes never reached main. This PR carries the same branch onto main; its diff against main is exactly #63's changes.

🤖 Generated with Claude Code

michaellchung and others added 3 commits September 25, 2026 05:32
Megatron's argument validation refuses tensor or context parallelism
without it.  The SSH harness already sets it per container; the sky/Modal
island task did not, so the first TP4 island (Qwen3.6-27B on 8xH100)
failed in parse_args before loading anything.
The island envs forced the unfused attention backend.  The learner already
passes --attention-backend per recipe and Megatron sets the three NVTE
variables from that flag in every actor, so the pin was redundant for the
generic provider and fatal for the flash recipes: the Qwen3.6-27B island
asserted 'NVTE_FLASH_ATTN set to 0, but expected 1' at model construction.
The two commits on this branch had no OpenSpec change of their own.
Records them under the sky-rl-island-runtime capability that
fix-sky-island-ray-stop introduces: the island's env must satisfy
Megatron's precondition for TP>1 and must not pin an attention backend
the learner did not choose.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@michaellchung
michaellchung merged commit e21a7ff into main Sep 29, 2026
1 of 6 checks passed
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