Skip to content

docs(m11): refresh stale deferred-work statuses (F19/F20 save bracket, LOW1)#37

Open
JunzheJoe wants to merge 1 commit into
zhenyu/miles-mvp-e2efrom
joe/docs-guide-status-refresh
Open

docs(m11): refresh stale deferred-work statuses (F19/F20 save bracket, LOW1)#37
JunzheJoe wants to merge 1 commit into
zhenyu/miles-mvp-e2efrom
joe/docs-guide-status-refresh

Conversation

@JunzheJoe

Copy link
Copy Markdown
Collaborator

Two factual status refreshes found during the F1-F12 code review — no code changes, no decisions needed:

  1. F19/F20 (saving) — §6.1 said Unchanged, but the exact "real fix" the historical row prescribes (explicit onload → save_model → offload bracket) is already implemented in miles/utils/rlix_train_loop.py:203-218. Marked the save side resolved; eval-at-final-rollout stays open.

  2. LOW1generate_rollout_fully_async declares and uses the rlix_hooks kwarg (examples/fully_async/fully_async_rollout.py:225, begin/bump/end at L275/385/422), so call_rollout_fn's inspect.signature forward cannot silently no-op on the legacy path. Marked verified-closed in both the §3.5 follow-ups table and the §6.1 entry, with a pointer to C24 (feat(rlix-validation): C24 — fail fast on RLix + experimental rollout refactor miles#33) for the experimental-refactor path.

Historical table rows are preserved verbatim per the doc's own append-only convention; updates carry dated markers.

…, LOW1)

Two status rows verified during the F1-F12 code review:

- F19/F20 (saving): the 'real fix' the historical row calls for —
  explicit onload -> save_model -> offload bracket — is already
  implemented in miles/utils/rlix_train_loop.py:203-218. Save side
  resolved; eval-at-final-rollout remains open.
- LOW1: generate_rollout_fully_async does declare and use the
  rlix_hooks kwarg, so the inspect.signature forward cannot silently
  no-op on the legacy path. Closed; the experimental-refactor path
  (which does not thread hooks) is now rejected by C24 (rlops/miles#33).

Historical table rows are preserved verbatim per the doc's append-only
convention; updates land in the 6.1 addendum with dated markers.
@JunzheJoe JunzheJoe requested a review from zhenyulincs July 7, 2026 03:23
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