Skip to content

Validate retained model recipes on source-pinned OSMO inputs - #156

Merged
ElmoPA merged 1 commit into
mainfrom
codex/graph-integration-20260924/07-osmo-validation
Oct 7, 2026
Merged

ElmoPA merged 1 commit into
mainfrom
codex/graph-integration-20260924/07-osmo-validation

Conversation

@rpuns

@rpuns rpuns commented Sep 24, 2026

Copy link
Copy Markdown

Run the retained HPT and PI recipes through five real optimizer updates, validation, checkpoint saving, and the shared no-hardware inference loader on OSMO L40s. The matrix covers EVA, Aria, Mecka, Scale, retained HPT Yam, both cotraining paths, and a two-GPU HPT case. CPU input staging verifies complete disjoint episode manifests, public weights and tokenizer hashes before scheduling GPUs; outputs use conditional writes at new R2 prefixes.

PI exposes optional gradient checkpointing and sampler compilation controls while retaining existing defaults. This draft does not claim real-weight training, destination assembly, or legacy cutover has passed.

Validation: isolated 1,411 tests passed; 253/253 YAML compositions and constructor contexts passed; lock, Ruff check and format passed. The synthetic harness verifies five actual updates plus bound checkpoint reload. R2 probes confirmed both ordinary and multipart writes reject replacing an existing object. The full snapshot and log hashes are recorded in docs/integration/evidence/osmo-harness-cpu-receipt.json.

Stacked on #155. GPU receipts will be added after execution; DQC remains stopped.

ElmoPA commented Oct 7, 2026 •

Copy link
Copy Markdown
Collaborator

Merge activity

  • Oct 7, 9:07 PM UTC: A user started a stack merge that includes this pull request via Graphite.
  • Oct 7, 9:19 PM UTC: Graphite rebased this pull request as part of a merge.
  • Oct 7, 9:19 PM UTC: @ElmoPA merged this pull request with Graphite.

@ElmoPA
ElmoPA changed the base branch from codex/graph-integration-20260924/06-ci-gates to graphite-base/156 October 7, 2026 21:16
@ElmoPA
ElmoPA changed the base branch from graphite-base/156 to main October 7, 2026 21:17
@ElmoPA
ElmoPA force-pushed the codex/graph-integration-20260924/07-osmo-validation branch from 37cb8db to a9df3c1 Compare October 7, 2026 21:18
@ElmoPA
ElmoPA merged commit 1205811 into main Oct 7, 2026
1 of 4 checks passed
ElmoPA added a commit that referenced this pull request Oct 7, 2026
Retained recipes could silently accept the wrong data frame, and network-only preflight missed broken language schedulers. Add model-owned preprocessing requirements to the restored root recipes and enforce them before model construction and during checkpoint/inference binding. Shared runtime code compares declared data state without model-family dispatch.

Execute the pinned EgoVerse source to preserve ten numerical data/normalization/metric references and both full language learning-rate schedules. Preserve camera-frame MSE at rotation wraparound, sparse-video upload FPS and tempo video lifecycle. Extend offline constructor checks through the actual model wrapper, training behavior, optimizer and scheduler; repair the missing UNITE baseline identity declaration without changing its objective.

Validation: isolated snapshot `813e786f7e50e7dfae897e1bc55cc4130269858c`, **1,430 CPU tests passed**; locked dependencies and pinned Ruff pass; **254/254** YAML and component contexts pass. Fixture provenance and tolerances are documented in `docs/integration/FIXED_FIXTURE_PARITY.md`.

This draft is layer 08, based on #156. Real-weight/data OSMO execution is blocked by HTTP 403 at the service/login endpoints; no integration GPU workflow has been submitted. Nested deployment contracts and assembled-destination validation remain open. No legacy-runtime cutover or deletion is authorized by these CPU results; DQC remains stopped.
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