Skip to content

feat(releases): qualify strata 0.2.16 with omp 19.0.0 on three hosts and the studio routes - #104

Draft
alphastorm wants to merge 25 commits into
mainfrom
feat/strata-0216-omp-1900
Draft

alphastorm wants to merge 25 commits into
mainfrom
feat/strata-0216-omp-1900

Conversation

@alphastorm

Copy link
Copy Markdown
Owner

Qualifies stock Strata v0.2.16 (engine 0.1.62) with stock OMP v19.0.0 on the three hosts and the Studio routes.

Draft while the host windows run. So far:

  • Admit stock v0.2.16's terminal/ layout behind a Strata-tag gate (v0.2.13/v0.2.12 rollback roots keep their call plans and identities; contract fixtures from 75d01fa).
  • Admit engine 0.1.62: reviewed environment groups (the stock 0.1.62 sm_86 table drops STRATA_ADAPT_NOWAIT), start-path sources across serve/server_parts/, the per-slot DFlash2 server line and the draft-vocabulary line.
  • Pin undeclared_tool_calls: "call" on every new profile (owner decision of 2026-11-12) through stock write_config.
  • Admit stock OMP 19.0.0: version-gated rendered settings (no composer predictions, no named bash services, no npm version fetch), measured G13/G18 compaction thresholds for the v19 fixed-context gate.
  • Draft PRO df2 and df2tuned, RTX 5090 262K envg and 131K, pair df2conf05 and both Studio routes; open the v0.2.16 + OMP 19.0.0 campaign.

Qualification evidence, receipts and docs follow after the host windows.

- Keep rollback renders intact; disable v19 service launch, composer and spelling predictions, and the Codex registry refresh.
- Ignore nonterminal RPC end events. Probe the versioned expert-heat route and force G13 compaction above the measured fixed context.
- Set G13/G18 to 10,000 after bracketing the 5,940-token fixed-context boundary; reject dead ends and surviving daemon workers.
- Preserve 166 rollback renders and frozen G24/G25 identities; scope new waits and terminal accounting to the unfrozen clients.
- Every committed campaign validates as it is; only schema 1 and 2 records must refuse OMP fields.
- The first schema 3 tuple campaign (v0.2.16 + OMP 19.0.0) is committed next.
- Open the v0.2.16 + OMP 19.0.0 tuple campaign (schema 3): PRO df2, RTX 5090 131K and 262K, pair df2conf05, Studio df2 route.
- Draft PRO df2 and df2tuned, RTX 5090 262K envg and 131K, and pair df2conf05 from their v0.2.13 predecessors: engine 0.1.62 component, no llama.cpp archive, undeclared_tool_calls call, stock 0.1.62 sm_86 table on the pair.
- Draft both Studio routes against the new PRO server fingerprints.
- Add the drafts to the CI suite matrix and render COMPATIBILITY.
- Check both resumed tool cycles in one conversation; identify the v19 sibling journal and retain the single-file assertion on v18.
- Exercise DeepSeek G18 budgets with an OmpRun that supplies the actual event-list contract.
- Keep pip index resolution disabled and preserve the pre-setup engine and MTP guards.
- Exercise stock v0.2.16 pip_install without a prior stamp against the Windows runtime and monitor lock.
…flag

- A flag is a whole whitespace-delimited token and its value runs to the next flag; the old split consumed the space before --draft after the valueless --trim-stage-weights, so the pair's first v0.2.16 start refused "exactly one --draft".
- Regression with the real pair argv shape.
- Stock v0.2.16 ordered_gpus keeps the configured order with no message under --gpu-roles; the automatic rule moved to gpu_roles.speed_stage_order unchanged.
- The pair's df2conf05 measurement start (2026-11-19) parsed complete, kept_configured, staged [1, 0], layers 0-23 / 24-47, one engine generation.

This branch has not been deployed

No deployments
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