Skip to content

Upgrade llama.cpp from b9975 to b9982 and drop merged patches#339

Merged
bernardladenthin merged 3 commits into
mainfrom
claude/patch-review-b9981-b9982-4rafas
Jul 13, 2026
Merged

Upgrade llama.cpp from b9975 to b9982 and drop merged patches#339
bernardladenthin merged 3 commits into
mainfrom
claude/patch-review-b9981-b9982-4rafas

Conversation

@bernardladenthin

Copy link
Copy Markdown
Owner

Summary

  • Upgrade pinned llama.cpp version from b9975 to b9982 across CMake, Java constants, and documentation
  • Remove two local patches that have been merged upstream:
    • 0004-pr23116-server-per-request-reasoning-budget-tokens.patch (merged in upstream PR #23116)
    • 0005-server-recurrent-near-prompt-end-checkpoints.patch (merged upstream)
  • Update patch documentation in CLAUDE.md to reflect the removal and current patch status

Test plan

  • CI is green on this branch
  • Docs updated (version references in CLAUDE.md, README.md, LlamaCppVersion.java, CMakeLists.txt, and breaking-changes history)

Related issues / PRs

Upstream merges:

  • ggml-org/llama.cpp#23116 — per-request reasoning budget tokens support
  • Recurrent model checkpoint optimization (merged upstream)

Checklist

  • I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
  • My commits follow Conventional Commits
  • No security-sensitive changes

https://claude.ai/code/session_01KYQqYWhZefKd5mus7FGM9W

claude added 3 commits July 13, 2026 07:10
Drop patch 0005 (recurrent near-prompt-end checkpoints): upstream's own
server-context.cpp create_checkpoint/do_checkpoint now solve the same
checkpoint-starvation problem with a strict superset fix (id_task-tracked
eviction + an unconditional near_prompt_end min-step exemption, vs our
patch's recurrent/hybrid-only scoping), so 0005 no longer applies and is
redundant. Verified via a fresh fail-loud cmake configure against b9981
with the remaining seven patches (0001-0004, 0006-0008) applying cleanly.
Update the b9975-b9981 verification row with the real cmake --build +
ctest results now that the full native build finished cleanly.
Drop patch 0004 (PR#23116 carry): upstream merged it verbatim in this
release — server-common.cpp's oaicompat_chat_params_parse gets the same
reasoning_budget_tokens/reasoning_budget_message per-request precedence
change, plus the same two tests/test-chat.cpp regression cases, so the
local patch is now redundant. A fresh native build + full ctest run is
in progress to confirm the remaining patches still apply and compile
cleanly against b9982; results will follow in a subsequent commit.
@bernardladenthin
bernardladenthin merged commit 9ac1dff into main Jul 13, 2026
10 of 14 checks passed
@bernardladenthin
bernardladenthin deleted the claude/patch-review-b9981-b9982-4rafas branch July 13, 2026 07:24
@sonarqubecloud

Copy link
Copy Markdown

bernardladenthin pushed a commit that referenced this pull request Jul 13, 2026
…ry rows

Additive-only bump (SYCL backend internals + auto-followed WebUI only,
no patch-target file touched); all six remaining patches apply
unchanged. Full native build + ctest: 485/485 passing.

Also recovers the b9981-b9982 history-doc rows from orphaned commit
6eacc95, which never made it into main - PR #339 appears to have been
merged at an earlier commit on that branch before the follow-up commit
landed.
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