Skip to content

Account for grouped checkpoint, head, and enclosing forward buffers - #898

Merged
bradhilton merged 42 commits into
mainfrom
schulman/848-checkpoint-head-demand-20260915
Sep 24, 2026
Merged

bradhilton merged 42 commits into
mainfrom
schulman/848-checkpoint-head-demand-20260915

Conversation

@bradhilton

@bradhilton bradhilton commented Sep 15, 2026 •

Copy link
Copy Markdown
Collaborator

Account for identified grouped checkpoint, head and enclosing-forward buffers in memory admission. Carry the shape-aware demand through plan selection, splitting, recovery and diagnostics; retain empirical floors and mark replay incomplete when immutable runtime facts are unavailable.

Merged with Brad’s explicit approval as a655c051, after #888 landed as e2e8969f. An ancestry-only successor 25bff2e9 preserved the complete reviewed 4a519f6d tree. The final merge also preserves concurrently landed #945; three independent composition reviews cleared exact tree 1a015493, and 75 focused report/evidence tests passed against that combination. Relative to #888, the physical-free measurement, reserve and cache-release policy are unchanged. The new demand estimates can change which plans are accepted, split or refused. No public API or art.megatron change is included.

Additional independent review: Fable cleared this exact #898 diff with explicit behavior limits. It adds partial memory components and reprices selected slots; admission, splitting and execution order can change in either direction across different profiles/plans. Additional metadata walks and exact-plan construction have unmeasured latency. Grouped planner-report replay remains incomplete where immutable runtime facts are missing. Brad explicitly accepted both #888’s recovery behavior and #898’s admission behavior after reviewing these limits.

The completed combined native run used source f80dba32: the full eight-pair prefix returned six backwards and one optimizer update, followed by the unchanged flat16 target—16 requests, 101,280 packed/logical rows—and its backward. Seven output windows and 660 parameter entries were retained. Original native/container exits were 0; the four owned remote UIDs and final known host union were closed. Root accepted source-bound completion and owned cleanup in B/combined-full8-native-v6-final-root-acceptance.json (47cb19ea), joining independent review 72f896c5, strict result b18a78f0 and final closure ae3f0e39. Here B = /home/brad/.local/share/schulman/art848-resume-20260921-root.

Two additional single-H200 workloads have now completed on f80dba32:

  • The full hidden-output probe completed 256 items / 384 requests / 5,622,779 logical tokens through one API call, with the planner choosing 27 internal waves. The caller did not manually split the input, and original output retention/pooling produced finite 256 × 2048 features. Diagnostic capture remains incomplete because the private reader rejected a source hardlink; no admission-accuracy or memory-peak claim follows. Receipt: B/052-local-forward-success-20260923/acceptance.json (b9f155f3).
  • A two-group, four-alternative DPO-shaped callback completed discriminator-gradient accumulation, mixed policy/reference passes and one finite joint optimizer update. This is a structural surrogate, not the exact historical failed step. Both original caller cache-release sites remain, so it does not qualify automatic recovery budgeting. Receipt: B/053-local-structural-success-20260923/acceptance.json (a90b2033).

Both original tasks and managed processes returned 0, and known owned resources are closed. Independent readout reviews cleared completion within those limits. These runs exclude the later #931 workspace allowances; the same-source #888 ON/OFF comparison also completed. Its strict result is INCONCLUSIVE because execution layouts and reserved-cache state differed; separate full-work timings were 518.711 s enabled versus 504.801 s disabled (+2.756%), descriptive only. Exact readout and limits are in #888 and B/888-current3b0-onoff-result-20260923/.

The landing tree 06fc8182 has all 14 planner modules, 37 related test files and CI routing byte-identical to f80d. It additionally preserves current main's Tinker/Responses/docs changes; the two upstream modules already passed 32 tests. This is a source-equivalence mapping, not a claim that the whole 4a519 package was executed.

Supporting #888 evidence remains source-scoped: one positive pressure release followed by backward/update (B/888-current674-pressure-final-root-acceptance.json), grouped save with exact fresh-state reload and a further update (B/888-current674-grouped-save-fresh-load-root-acceptance.json), and the fixed 12-update observation of +0.21668% (B/888-current-device-timing-observation-root-acceptance.json). The timing pair had no explicit releases and strict warmed status UNKNOWN; it is not a sustained overhead bound.

Focused planner/report checks passed 98 tests; the combined recovery/evidence fixture checks passed 15 cases + 2 subtests. The exact successor's required default prek ran once: Ruff, formatting and lock checks passed; ty retained the same 11 diagnostics in unchanged files, with no new diagnostic or byte change. At exact head 4a519f6d, quality CI and two-H200 validation both passed; Package Install was skipped. Three independent source reviewers cleared this exact composition. The local baseline type-check failure remains preserved.

The strict full8 reader reports source-bound completion only. Counter continuity/admission qualification, numerical comparisons and an old-runtime counterfactual remain unavailable; historical RNG, allocator and dataloader state were not recreated. Prior refusal, setup/readout and numerical failures remain preserved in /home/brad/.local/share/schulman/pr888-pr898-evidence-draft-20260923-cold-v1/ARTIFACTS.md. This closes the specific missing prefix-update/flat16 completion observation, not generic #848/#870, gradient parity, all-shape memory safety or a universal 5% overhead claim. Brad approved landing this admission change; the broader qualification limits remain.

Final successor 25bff2e9 passed quality CI and two-H200 validation. Root verified the actual merge parent, complete tree, file modes and patch against the reviewed combination. Post-merge main CI and image publication are separate automatic follow-up checks; no manual deployment was performed. Exact receipts: B/merge888898-20260923/.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 15, 2026 08:08 — with GitHub Actions Active

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 15, 2026 08:26 — with GitHub Actions Active

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 15, 2026 08:52 — with GitHub Actions Active

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton force-pushed the schulman/848-checkpoint-head-demand-20260915 branch from bec2794 to b430f5e Compare September 16, 2026 10:03
@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 16, 2026 10:04 — with GitHub Actions Active
@bradhilton bradhilton changed the title Account for checkpoint retention and head backward in TrainerRank admission Account for grouped checkpoint and shared-expert memory in TrainerRank admission Sep 16, 2026

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 16, 2026 10:22 — with GitHub Actions Active

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 16, 2026 10:43 — with GitHub Actions Active

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 16, 2026 11:16 — with GitHub Actions Active

@bradhilton bradhilton left a comment •

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Superseded review text removed on 2026-09-17; see the consolidated review-history comment on this PR. Current-head reviews were left in place.

@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 22, 2026 08:38 — with GitHub Actions Active
@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 23, 2026 01:16 — with GitHub Actions Active
@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 23, 2026 01:45 — with GitHub Actions Active
Apply ART942 (3518701) to the private ART898 candidate 61ef82a. Preserve the checkpoint-aware snapshot and replay fields in the extracted helper, support report formats 1 and 2, and retain both diagnostic source fingerprints.

Validation: source AST comparisons, compile-only checks, and git diff --check. Framework imports, runtime tests, GPU qualifications, and publication were not performed.
@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 23, 2026 15:23 — with GitHub Actions Active
Keep the PR888 recovery episode and actual forward/backward work and observer
cost operands when adopting main's planner diagnostics. Preserve unavailable
handoff memory fields and keep the PR898 cost model out of this merge.

Validation: 15 focused cases and 2 subtests passed on this exact tree.
Required default prek ran once: Ruff, format and lock checks passed; ty failed
with the same 11 retained diagnostics and no new diagnostics. No hook edits.
Compose the existing backward recovery and planner diagnostics with the
checkpoint, head, and enclosing-buffer estimates. Planner runtime and
related tests exactly match the tested private f80d composition; retain
current main Tinker and Responses changes.

Required default prek was run once before commit: Ruff, formatting, and
lock checks passed; ty retains the same 11 diagnostics in unchanged files.
No new diagnostics or hook byte changes. No native or CPU test replay.
@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 23, 2026 20:06 — with GitHub Actions Active
@bradhilton
bradhilton marked this pull request as ready for review September 23, 2026 20:32
@bradhilton
bradhilton deployed to trainer-rank-gpu-validation September 23, 2026 20:32 — with GitHub Actions Active
bradhilton added a commit that referenced this pull request Sep 23, 2026
Preserve the current #888/#898 planner integration while carrying the existing
7D eager-statistics and 3D no-grad logits capacity allowances, head tests and
CI routing. Keep public #931 ancestry and the reviewed current #898 base.
Preserve the exact reviewed 4a519f6 tree while joining main after the squash merge of PR888. No file or runtime changes.

This branch was successfully deployed

1 active deployment
trainer-rank-gpu-validation — 25bff2e9 Deployed Sep 24, 2026 by bradhilton via Run on 2x H200 #633
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