Muse Glimmer 30B on llama.cpp (macOS)

Convenience Makefile for downloading and serving Meta Muse Glimmer 30B with
llama.cpp on Apple Silicon.

Setup follows
Run Muse Glimmer-30B Locally with llama.cpp on macOS,
which is the reference for the build flags, the three GGUFs, and the server
invocation wrapped up here. Read it first if you want the reasoning behind any
of the settings; this repo just automates them.

llama.cpp lives in a separate folder

This repo contains no llama.cpp source and no build tree — only the
Makefile. The upstream clone is a sibling directory, which keeps these local
targets out of a tree you'll be pulling and rebuilding:

~/projects/llama-cpp/
├── llama.cpp/                  <- upstream clone (source + build/)
└── llama-cpp-muse-glimmer/     <- this repo (Makefile only)

The clone's location is the LLAMA_CPP variable at the top of the Makefile,
and everything else derives from it:

LLAMA_CPP ?= $(HOME)/projects/llama-cpp/llama.cpp
BUILD_DIR ?= $(LLAMA_CPP)/build

SERVER ?= $(BUILD_DIR)/bin/llama-server
CLI    ?= $(BUILD_DIR)/bin/llama-cli
LLAMA  ?= $(BUILD_DIR)/bin/llama

Point LLAMA_CPP at your own clone — edit that line, or override it per
invocation:

make build LLAMA_CPP=~/src/llama.cpp

make build passes the source directory explicitly (cmake -S $(LLAMA_CPP) -B $(BUILD_DIR)), so nothing is built relative to the current directory.

You need a recent llama.cpp build

Muse Glimmer is a new architecture, so an older llama.cpp cannot load the
model at all — it fails on the unknown arch rather than doing anything
graceful.

• Build b10353 or newer is required. The article notes Homebrew's formula
  was still on b10330, so brew install llama.cpp is not sufficient — build
  from source.

• An existing clone needs updating. If you cloned llama.cpp before Muse
  Glimmer landed, make build compiles a tree that can't load the model. Pull
  first:

  git -C ~/projects/llama-cpp/llama.cpp pull
  

• make build verifies this for you. It depends on check-arch, which
  greps $(LLAMA_CPP)/src/llama-arch.cpp for LLM_ARCH_MUSE_GLIMMER and stops
  before CMake runs if the symbol isn't there, so a stale checkout fails in a
  secon