Add Apple Silicon (Metal) support

.github/workflows/ci.yml

@@ -13,7 +13,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        os: [ubuntu-latest, windows-latest]
+        os: [ubuntu-latest, windows-latest, macos-latest]
         python-version: ["3.11", "3.12", "3.13"]
       fail-fast: false
 

README.md

@@ -1,6 +1,6 @@
 # 🏎️ Dyno — llama.cpp Auto-Tuner & Benchmark
 
-**Dyno** is an open-source CLI that auto-tunes and benchmarks [llama.cpp](https://github.com/ggml-org/llama.cpp) / [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp) inference on NVIDIA GPUs, producing reproducible, shareable results.
+**Dyno** is an open-source CLI that auto-tunes and benchmarks [llama.cpp](https://github.com/ggml-org/llama.cpp) / [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp) inference on NVIDIA and Apple Silicon GPUs, producing reproducible, shareable results.
 
 ```bash
 pipx install llama-dyno
@@ -34,8 +34,8 @@ dyno report ~/Downloads/my-model.q4_k_m.gguf
 ## Prerequisites
 
 - **Python 3.11+**
-- **NVIDIA GPU** with drivers + CUDA
-- **llama-bench** binary from [llama.cpp](https://github.com/ggml-org/llama.cpp) or [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp)
+- **A GPU**: NVIDIA (drivers + CUDA) or Apple Silicon (M-series, uses Metal + unified memory)
+- **llama-bench** binary from [llama.cpp](https://github.com/ggml-org/llama.cpp) or [ik_llama.cpp](https://github.com/ikawrakow/ik_llama.cpp) — built with the matching backend (CUDA or Metal)
 
 Install llama.cpp:
 
@@ -182,7 +182,7 @@ pytest
 
 - GUI
 - Multi-GPU
-- Non-NVIDIA backends (AMD, Intel, Apple Silicon)
+- AMD (ROCm) / Intel backends (planned)
 - Server hosting for results
 
 ## License

pyproject.toml

@@ -5,14 +5,14 @@ build-backend = "hatchling.build"
 [project]
 name = "llama-dyno"
 version = "1.1.0"
-description = "Auto-tune and benchmark llama.cpp / ik_llama.cpp inference on NVIDIA GPUs"
+description = "Auto-tune and benchmark llama.cpp / ik_llama.cpp inference on NVIDIA and Apple Silicon GPUs"
 readme = "README.md"
 requires-python = ">=3.11"
 license = "Apache-2.0"
 authors = [
   { name = "Flowdesk Systems", email = "lachy@flowdesk.systems" },
 ]
-keywords = ["llama.cpp", "benchmark", "tuning", "gpu", "nvidia", "inference"]
+keywords = ["llama.cpp", "benchmark", "tuning", "gpu", "nvidia", "apple-silicon", "metal", "inference"]
 classifiers = [
   "Development Status :: 3 - Alpha",
   "Environment :: Console",

src/llama_dyno/cli.py

@@ -206,7 +206,8 @@ def detect():
     table.add_column("Value")
 
     table.add_row("GPU", hw.gpu_name)
-    table.add_row("VRAM", f"{hw.vram_total_mib} MiB")
+    mem_label = "Unified memory" if hw.gpu_name.startswith("Apple") else "VRAM"
+    table.add_row(mem_label, f"{hw.vram_total_mib} MiB")
     table.add_row("Driver", hw.driver_version)
     table.add_row("CUDA", hw.cuda_version or "N/A")
     table.add_row("CPU", hw.cpu_name)

src/llama_dyno/detect.py

@@ -151,6 +151,28 @@ def _detect_ram() -> int:
     return 0
 
 
+def _apple_silicon_gpu() -> tuple[str, int] | None:
+    """Detect an Apple Silicon GPU as (name, usable memory MiB), else None.
+
+    Apple GPUs share unified memory with the CPU — there is no discrete VRAM, so
+    we report total RAM as the pool the GPU can draw from (macOS lets Metal use
+    most of it).
+    # ponytail: unified mem ≈ total RAM; refine via Metal recommendedMaxWorkingSetSize if OOM heuristics misfire
+    """
+    if platform.system() != "Darwin" or platform.machine() != "arm64":
+        return None
+    chip = ""
+    try:
+        out = subprocess.run(
+            ["sysctl", "-n", "machdep.cpu.brand_string"],
+            capture_output=True, text=True, timeout=5,
+        )
+        chip = out.stdout.strip()
+    except Exception:
+        pass
+    return chip or "Apple Silicon GPU", _detect_ram()
+
+
 def _find_binary(name: str) -> str | None:
     """Find a binary in PATH. Returns path or None."""
     # WSL detection: prefer Linux binaries; Windows binaries under /mnt/c/
@@ -245,6 +267,13 @@ def detect_ik_features(binary_path: str | None = None) -> IkFeatures:
 def detect_hardware() -> HardwareFingerprint:
     """Fingerprint the current hardware and detect llama.cpp backend."""
     gpu_name, vram, driver, cuda_ver = _detect_gpu()
+    # No NVIDIA GPU found — fall back to Apple Silicon (unified memory) if present.
+    if gpu_name == "Unknown" or vram == 0:
+        apple = _apple_silicon_gpu()
+        if apple:
+            gpu_name, vram = apple
+            driver = f"macOS {platform.mac_ver()[0]}"
+            cuda_ver = None
     cpu_name, cpu_cores = _detect_cpu()
     ram = _detect_ram()
 

src/llama_dyno/tune.py

@@ -397,9 +397,9 @@ def _hill_climb(
 
 
 def _detect_vram_mib() -> int:
-    """Detect total GPU VRAM in MiB (pynvml, then nvidia-smi). 0 if unknown.
+    """Detect GPU memory in MiB (NVIDIA VRAM, then Apple unified memory). 0 if unknown.
 
-    A single monkeypatchable seam for the tuner's VRAM heuristic.
+    A single monkeypatchable seam for the tuner's memory heuristic.
     """
     try:
         import pynvml
@@ -416,9 +416,13 @@ def _detect_vram_mib() -> int:
             ["nvidia-smi", "--query-gpu=memory.total", "--format=csv,noheader,nounits"],
             capture_output=True, text=True, timeout=5,
         )
-        return int(out.stdout.strip()) if out.stdout.strip() else 0
+        if out.stdout.strip():
+            return int(out.stdout.strip())
     except Exception:
-        return 0
+        pass
+    from .detect import _apple_silicon_gpu
+    apple = _apple_silicon_gpu()
+    return apple[1] if apple else 0
 
 
 def _estimate_model_size(model_path: str) -> int:

tests/test_detect.py

@@ -0,0 +1,62 @@
+"""Tests for hardware detection, focused on the Apple Silicon path (mocked)."""
+
+from __future__ import annotations
+
+import subprocess
+
+from llama_dyno import detect
+
+
+def _fake_sysctl(out: str):
+    def run(cmd, capture_output=True, text=True, timeout=5, **kw):
+        return subprocess.CompletedProcess(cmd, 0, stdout=out, stderr="")
+    return run
+
+
+def test_apple_silicon_gpu_on_arm_mac(monkeypatch):
+    monkeypatch.setattr(detect.platform, "system", lambda: "Darwin")
+    monkeypatch.setattr(detect.platform, "machine", lambda: "arm64")
+    monkeypatch.setattr(detect, "_detect_ram", lambda: 16384)
+    monkeypatch.setattr(detect.subprocess, "run", _fake_sysctl("Apple M2 Max\n"))
+
+    assert detect._apple_silicon_gpu() == ("Apple M2 Max", 16384)
+
+
+def test_apple_silicon_gpu_none_on_linux(monkeypatch):
+    monkeypatch.setattr(detect.platform, "system", lambda: "Linux")
+    monkeypatch.setattr(detect.platform, "machine", lambda: "x86_64")
+    assert detect._apple_silicon_gpu() is None
+
+
+def test_apple_silicon_gpu_none_on_intel_mac(monkeypatch):
+    monkeypatch.setattr(detect.platform, "system", lambda: "Darwin")
+    monkeypatch.setattr(detect.platform, "machine", lambda: "x86_64")
+    assert detect._apple_silicon_gpu() is None
+
+
+def test_detect_hardware_uses_apple_fallback(monkeypatch):
+    monkeypatch.setattr(detect, "_detect_gpu", lambda: ("Unknown", 0, "unknown", None))
+    monkeypatch.setattr(detect, "_apple_silicon_gpu", lambda: ("Apple M2 Max", 32768))
+    monkeypatch.setattr(detect.platform, "mac_ver", lambda: ("14.5", ("", "", ""), "arm64"))
+    monkeypatch.setattr(detect, "_find_binary", lambda name: None)
+
+    hw = detect.detect_hardware()
+    assert hw.gpu_name == "Apple M2 Max"
+    assert hw.vram_total_mib == 32768
+    assert hw.cuda_version is None
+    assert hw.driver_version.startswith("macOS")
+
+
+def test_detect_hardware_keeps_nvidia(monkeypatch):
+    monkeypatch.setattr(detect, "_detect_gpu", lambda: ("NVIDIA RTX 4070", 12282, "550.0", "12.4"))
+
+    def _boom():
+        raise AssertionError("Apple fallback must not run when NVIDIA is present")
+
+    monkeypatch.setattr(detect, "_apple_silicon_gpu", _boom)
+    monkeypatch.setattr(detect, "_find_binary", lambda name: None)
+
+    hw = detect.detect_hardware()
+    assert hw.gpu_name == "NVIDIA RTX 4070"
+    assert hw.vram_total_mib == 12282
+    assert hw.cuda_version == "12.4"