Skip to content

Bump llama.cpp to 785322d17 (b12032+1), release v0.8.66 - #121

Merged
nyo16 merged 1 commit into
masterfrom
bump-llama-785322d17
Nov 1, 2026
Merged

nyo16 merged 1 commit into
masterfrom
bump-llama-785322d17

Conversation

@nyo16

@nyo16 nyo16 commented Nov 1, 2026

Copy link
Copy Markdown
Owner

llama.cpp 418f905a4 → 785322d17 (b12032+1), 31 commits.

API: only common/common.h changed among the NIF's headers (fs_path_to_utf8 removed, not used). llama.h, ggml*.h, chat.h, speculative.h, llama-ext.h, llama-vocab.cpp and llama-context.cpp are untouched, and both *_default_params are unchanged. RPC protocol stays at 8.0.1.

Decision port: ports #30710 (server_decision_group_tasks caps the shared prefix at each question's first read position); "Ported at" is now 785322d17. scripts/decision_compare.exs against llama-server built from the same commit (Metal, -np 1) gives ALL MATCH on tinylaya, tinyopenjev, d1-2B, d1-3B, d1-omni-600M, vllm-decision2-Kai, vllm-d3-edge, Clef Flash and h2o-lightning-4b. The sharing column is unchanged.

Notable upstream changes:

  • recurrent seq_rm refuses rollbacks past the last ubatch's snapshots (#31368);
  • KV rotation fixes (#31073, #31406);
  • Gemma 3n/4 token_embd order (#31394);
  • Metal quantize kernels in fp math_mode(safe) (#31120).

Verified on M1 Max, Metal, residency sets on:

  • default 436;
  • release-guide set 588;
  • CI set 604;
  • embeddinggemma 447;
  • Clef 440;
  • Llama-3.2 :tools 446/447 (known);
  • mtp_sidecar 442 + 6 skipped;
  • CPU :tools 447;
  • CPU vm_exit 3;
  • :mtp_cancel is still the known race (3/4 abort, record updated);
  • format, credo and dialyzer clean;
  • the Hex source build clones 785322d17 and compiles.

Upstream defects #1–#3 are unchanged (re-check added to docs/release-guide.md).

31 commits from 418f905a4 (b11986+16) to b12032+1. Ports #30710 into
decision.cpp: a shared decision prefix also stops at each question's first
read position. No NIF change was needed; the only header change removes
fs_path_to_utf8, which the NIF does not call. decision_compare.exs: ALL
MATCH on 9 models against llama-server at the same commit.
@nyo16
nyo16 merged commit b57e5d0 into master Nov 1, 2026
9 checks passed
@nyo16
nyo16 deleted the bump-llama-785322d17 branch November 1, 2026 21:37
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