Skip to content

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

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
JunzheJoe merged 1 commit into from
Jul 19, 2026
Merged

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

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions docs/m11-implementation-guide.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -536,7 +536,7 @@ Four small fixes from the M11 review report.
| MED | F2 | configurable free-mem threshold via `MILES_MAX_RESIDUAL_GPU_MEM_GB` | howard989's PRs `rlops/rlix#11` + `rlops/miles#3` in flight (v2 per @taoluo) |
| MED | MED1 | concurrent-resize stress test under `max_concurrency=4` | Open (Codex carryforward) |
| NOTE | F8 / F10 | `orchestrator.cleanup_stale_pipelines()` RPC | M11.3 production hardening |
| LOW | LOW1 | verify `generate_rollout_fully_async` accepts `rlix_hooks` kw | ~30 min check |
| LOW | LOW1 | verify `generate_rollout_fully_async` accepts `rlix_hooks` kw | ✅ Verified closed _(2026-07-06, F1-F12 review)_ — see §6.1 |

## §3.6 Tianye's PR #14 — finalize-always-continue + release-on-sync-failure ✅ MERGED 2026-05-18 _(added 2026-05-24)_

Expand Down Expand Up @@ -698,7 +698,7 @@ The table above is the historical M11.1 / M11.2 Option A deferred-work record an
| **F22 (relaxed)** | **Partially resolved by Phase 3 Option β** (env-gated `MILES_INIT_DEFER_ADD_WORKER=1` lands engines in `loading` → `finish_init_offload` → `offloaded`). Full F22 architectural rebuild still deferred to M11.5; Option β provides the same observable Gate 4(c) semantics. | _(added 2026-05-24)_ |
| **M11.3** | M11.2 real overlap (2 pipelines on shared GPUs) now validated; 3+ still deferred. | _(added 2026-05-24)_ |
| **F95.1** | Still deferred; real overlap M11.2 also unaffected because Option β's `finish_init_offload` parks engines before contention. | _(added 2026-05-24)_ |
| **F19/F20 (saving)** | Unchanged. | _(added 2026-05-24)_ |
| **F19/F20 (saving)** | **Save side resolved** _(updated 2026-07-06, F1-F12 review)_: `run_async_train_loop` wraps `save_model` in an explicit onload→save→offload bracket (`miles/utils/rlix_train_loop.py:203-218`), matching the "real fix" described in the historical row above. Eval-at-final-rollout remains open (still intentionally skipped in rlix-mode loops). | _(added 2026-05-24)_ |
| **Eval at final rollout** | Unchanged. | _(added 2026-05-24)_ |
| **Pytest coverage of M11 changes** | Tianye's PR#16 added 3 new test files (`test_orchestrator_death_is_benign.py`, `test_scheduler_apply_plan_invariants.py`, `test_gap_ratio.py` rewrite) covering planner durability + scheduler intent. Still no unit tests on `_wait_for_overlap_engines_offloaded` or `_split_pools_for_dual` (F9 behavior-tested via local Python invocation). | _(added 2026-05-24)_ |
| **R04-F1 (HIGH from review-report)** | **✅ RESOLVED** by Phase 1 (rlix `47bf02b` + miles `6513b25`). Verified on vast Attempt 0 via `MILES_INJECT_TRAIN_FAULT=1` injected smoke. | _(added 2026-05-24)_ |
Expand All @@ -719,7 +719,7 @@ New deferred-work entries surfaced during M11.2 real-overlap work (not in the hi
| **F7 (NOTE)** | `shrink_engines` pause_generation contract undocumented | Documentation gap | Reviewers confused by broad `except Exception` | **✅ RESOLVED** by miles `1487c3f` — docstring covers mode='retract', idempotency, 4xx vs 5xx, swallowing rationale; log includes engine_indices. | _(added 2026-05-24)_ |
| **F8 / F10 (NOTE)** | Detached coordinator persists after driver crash | M11.3 cleanup RPC scope | Stale actors on operator-killed runs | Still open; needs `orchestrator.cleanup_stale_pipelines()` RPC. | _(added 2026-05-24)_ |
| **F9 (NOTE)** | `_split_pools_for_dual` silently ignored extra GPUs on non-2N machines | Static-analysis finding | Silent GPU leak on 5/6/7-GPU machines | **✅ RESOLVED** by miles `1487c3f` — explicit `ValueError` when `num_gpus_per_node != 2*infer_pool_size`; error message points at MILES_DUAL_P* workaround. | _(added 2026-05-24)_ |
| **LOW1 (Codex carryforward)** | Verify `generate_rollout_fully_async` actually accepts `rlix_hooks` kw | `inspect.signature` forward silently no-ops if kw is missing | Demand signal path dead-codes silently | Open; ~30 min grep + startup assert. | _(added 2026-05-24)_ |
| **LOW1 (Codex carryforward)** | Verify `generate_rollout_fully_async` actually accepts `rlix_hooks` kw | `inspect.signature` forward silently no-ops if kw is missing | Demand signal path dead-codes silently | **✅ Verified closed** _(2026-07-06, F1-F12 review)_: `examples/fully_async/fully_async_rollout.py:225` declares `rlix_hooks: RLixHooks | None = None` and drives begin/bump/end (L275/385/422); the `inspect.signature` forward in `base_types.call_rollout_fn` matches. Note this covers the legacy call path only — the experimental-refactor path does not thread hooks and is now rejected at startup by C24 (`rlops/miles#33`). | _(added 2026-05-24)_ |

---

Expand Down