feat(pipeline): add device autodetect with mps-safe vram limit

ace_pipeline.py

@@ -0,0 +1,40 @@
+"""ACE-Step pipeline lifecycle: device autodetect, lazy load, cache mirror.
+
+Mirrors z-image-studio's `models.py` pattern. M0 only implements device
+detection — the pipeline class itself is filled in at M1.
+"""
+from __future__ import annotations
+
+
+def detect_device() -> str:
+    """Returns 'cuda', 'mps', or 'cpu' in priority order."""
+    try:
+        import torch  # local import: keep module import cheap for CI
+    except ImportError:
+        return "cpu"
+
+    if torch.cuda.is_available():
+        return "cuda"
+    # macOS: torch.backends.mps appeared in 2.0; guard for the rare absence
+    if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+        return "mps"
+    return "cpu"
+
+
+def vram_limit_for(device: str) -> int | None:
+    """Returns a VRAM cap in bytes for CUDA, None otherwise.
+
+    `torch.mps` has no `mem_get_info` — calling DiffSynth-style free-VRAM
+    gates with a numeric limit would crash on MPS. Returning None lets the
+    pipeline short-circuit those checks.
+    """
+    if device != "cuda":
+        return None
+    try:
+        import torch
+
+        free, _total = torch.cuda.mem_get_info()
+        # Leave 2 GiB headroom for activations
+        return max(0, free - 2 * 1024**3)
+    except Exception:
+        return None

tests/test_ace_pipeline.py

@@ -0,0 +1,19 @@
+"""L1 tests for device autodetect — no torch needed if we mock importlib."""
+from __future__ import annotations
+
+import ace_pipeline as ap
+
+
+def test_detect_device_returns_one_of_cuda_mps_cpu():
+    device = ap.detect_device()
+    assert device in {"cuda", "mps", "cpu"}
+
+
+def test_vram_limit_for_mps_is_none():
+    """MPS has no torch.mps.mem_get_info; return None so DiffSynth-style gates
+    short-circuit instead of crashing (z-image-studio paid this debug cycle)."""
+    assert ap.vram_limit_for("mps") is None
+
+
+def test_vram_limit_for_cpu_is_none():
+    assert ap.vram_limit_for("cpu") is None