Bump llama.cpp to 72b590d65 (b10989+1, +46 from b6b003d2c), release v0.8.51

[nyo16]

Bumps vendor/llama.cpp from b6b003d2c (b10944) to 72b590d65 — origin/master, one commit past b10989 (upstream v0.4.1). +46 commits. Pins
LLAMA_COMMIT in the Makefile to the same SHA and bumps @version to 0.8.51.

API compatibility

No diff in any header the NIF includes — llama.h, ggml-backend.h, chat.h, json-schema-to-grammar.h, speculative.h — and none in llama_model_default_params / llama_context_default_params.
c_src/ is untouched.

The one relevant header change is ggml-rpc.h: RPC_PROTO_MAJOR_VERSION 6 → 7 (ggml-org/llama.cpp#28789). SET_TENSOR gains a cache-flag byte, so a worker at this build refuses a v0.8.50 client and
vice versa — both ends of a two-node split must rebuild together.

The same upstream change fixes what motivated it: the RPC hash cache (:cache_dir on LlamaCppEx.RPC.Server) used to hash and persist every transfer over 10 MiB, including the activations
ggml_backend_sched copies between backends, filling the worker's disk during prefill on a split model. It now covers weight tensors only. The :cache_dir moduledoc is adjusted to say so.

Release-guide workaround #3 (ggml_backend_rpc_start_server returns void) is unchanged upstream; the preflight-bind workaround stays.

Notable upstream changes in range

• ggml-cpu: heap overflow in the rope work buffer from a CACHE_LINE_SIZE mismatch between C and C++ TUs; precompiled header disabled (#28882)
• common_speculative skips the target decode when llama_n_rs_seq says it is not needed (#28749)
• Maple 20B-A1B ternary MoE architecture, CPU only (#27000)
• Metal FA kernels for HSK=96 / HSV=64, MiniCPM3 (#28599)

Verification

macOS (Metal), M4 Max, LLAMA_BACKEND=metal source build:

Run	Result
mix test (no model)	430 passed, 157 excluded
--include smoke --include embeddings --include slow --include mtp (Qwen3.5-0.8B-UD-Q8_K_XL, Qwen3-Embedding-0.6B-f16, Qwen3.5-0.8B-MTP-Q8_0)	573 passed, 14 excluded
MTPE4BSidecarTest under --include mtp_sidecar (gemma-4-E4B-it-Q4_K_M + mtp-gemma-4-E4B-it-Q8_0)	6 passed
mix format --check-formatted	clean

Not re-run at this build: the Qwen 3.8 sidecar pair (MTPSidecarTest), rpc_live, dialyzer, and the Hex source-build round-trip. checksum.exs is regenerated by the precompile workflow after tagging.

Files

• vendor/llama.cpp — submodule → 72b590d65
• Makefile — LLAMA_COMMIT
• mix.exs — @version 0.8.50 → 0.8.51
• CHANGELOG.md — v0.8.51 section
• lib/llama_cpp_ex/rpc/server.ex — :cache_dir doc: "weight tensors"