Skip to content

Retire patch 0003 and the v0.7.0 qualification row - #185

Merged
leehack merged 1 commit into
mainfrom
retire-patch-0003-v070-row
Nov 8, 2026
Merged

leehack merged 1 commit into
mainfrom
retire-patch-0003-v070-row

Conversation

@leehack

@leehack leehack commented Nov 8, 2026

Copy link
Copy Markdown
Owner

Part of #145 (first bullet only; the SVE256 lane and #149 audit-note items are untouched).

What

  • Delete carried patch 0003 (embedding-mode scheduler reserve, Qualify and repair the v0.7.0 CPU crash after enabling context embeddings #114) and its series.json entry.
  • Drop the v0.7.0 row from LLAMADART_UPSTREAMS in validate_wrapper.yml.
  • Tests: ci_scope real-table assertion keeps only v0.8.0; expected manifest list drops 0003; the upstream_fix retirement test (incl. CRLF and missing-source cases) now runs on 0004; the Android marker-requirement test now uses 0005's libggml-cpu.so marker.
  • Docs: AGENTS.md, patches/llama.cpp/README.md, docs/v090_qualification.md, docs/v080_qualification.md updated; docs/v070_native_contracts.md notes the patch was retired.

Evidence (issue trigger: core shipped on v0.8.0)

llamadart core native pin
v0.17.1 - v0.17.5 v0.8.0-1
v0.18.0 v0.8.0-3
main v0.9.0

Per docs/v090_qualification.md, 0003 applied only to the v0.7.0 row (v0.8.0+ carry upstream #31086).

Kept

  • Patch 0004 and the v0.8.0 row (0004 is needed only by that row).

Validation

  • python3 -m unittest discover -s tests: 449 tests OK (1 skipped)
  • llama_cpp_patches.py manifest --upstream third_party/llama.cpp (v0.9.0): 0001, 0002, 0005, 0006

@leehack
leehack merged commit b998594 into main Nov 8, 2026
57 of 60 checks passed
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