Skip to content

Bump llama.cpp to 9daa0aa22, release v0.8.69 - #126

Merged
nyo16 merged 1 commit into
masterfrom
bump-llama-9daa0aa22
Nov 13, 2026
Merged

nyo16 merged 1 commit into
masterfrom
bump-llama-9daa0aa22

Conversation

@nyo16

@nyo16 nyo16 commented Nov 13, 2026 •

Copy link
Copy Markdown
Owner

Bumps llama.cpp from a691b2345 (b12205) to 9daa0aa22 (b12361+7), 163 commits. Upstream is now llama.cpp 1.0.0 / ggml 1.0.0. Version → v0.8.69.

NIF / API

  • The upstream prune (#30565, #29801) removes every deprecated llama.h function, llama_context_params.defrag_thold and several ggml ops. The NIF calls none of them. NextN/DFlash queries move from src/llama-ext.h to llama.h unchanged. llama_model_default_params is unchanged.
  • src/llama-vocab.cpp only drops deprecated wrappers (the tokenizer_st_partition and BOS/EOS rules are untouched), and tools/server/server-decision.* is untouched, so input_tokenize.cpp and the decision port need no change. decision.cpp is now "Ported at" 9daa0aa22.
  • RPC protocol 8.0.1 → 10.0.0 (#29413, #29801; GGML_OP_COUNT 102 → 99). Workers and clients must both run v0.8.69. The HELLO layout is unchanged, so rpc_probe_hello only moves its static_assert to 10.
  • #30415 stops upstream registration aborting on a dead endpoint. The NIF keeps its probe because upstream has no connect timeout and answers a memoized endpoint without contacting it. The RPC docs, docs/dgx-spark.md, docs/release-guide.md and the solution note are updated. rpc-add-server-no-abort is superseded; rpc-client-spin-wait now conflicts with #29413.
  • Upstream defects Add Jinja chat templates, penalty params, and OpenAI response format #1–Bump v0.4.4: update llama.cpp submodule to b8198 #3 in the release guide: re-checked as a source diff, all still present.

Verification (M1 Max, Metal)

  • mix test on a default build (442 passed) and on an LLAMA_RPC=1 build (442 passed), plus --include rpc_live against a local RPC.Server worker (443 passed).
  • :smoke :embeddings :slow :mtp :tools (607 passed), embeddinggemma-2 :embeddings, :decision, :decision_clef, :mtp_sidecar (Qwen3.8-27B) and test/vm_exit_test.exs: all green.
  • scripts/decision_compare.exs: ALL MATCH on 9 models against llama-server at the same commit.
  • :mtp_cancel (known broken): 4 of 7 runs abort with GGML_ASSERT(logits != nullptr). Recorded in test/mtp_model_test.exs.
  • mix format --check-formatted and mix dialyzer: clean.
  • Hex source build: mix hex.build, unpacked, LLAMA_BACKEND=cpu mix compile clones 9daa0aa22b8a… (= submodule) and builds.
  • Not run on a DGX Spark for this release.

163 commits from a691b2345 (b12205) to b12361+7; llama.cpp and ggml are
now 1.0.0. The upstream prune (#30565, #29801) removes no API the NIF
calls. decision.cpp is "Ported at" 9daa0aa22 with no hunk to port.
decision_compare.exs: ALL MATCH on 9 models against llama-server.

- RPC protocol 8.0.1 -> 10.0.0 (#29413, #29801): workers and clients must
  both run v0.8.69. The HELLO layout is unchanged, so rpc_probe_hello only
  moves its static_assert to 10.
- #30415 stops upstream registration aborting on a dead endpoint. The NIF
  keeps its probe (connect timeout, memoized endpoints); docs and the
  solution note say so. rpc-add-server-no-abort is superseded;
  rpc-client-spin-wait conflicts with #29413.
- Record :mtp_cancel at this commit (4/7 abort, logits != nullptr).
@nyo16
nyo16 merged commit 3f6c582 into master Nov 13, 2026
9 checks passed
@nyo16
nyo16 deleted the bump-llama-9daa0aa22 branch November 13, 2026 17:29
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