update finetuning pipeline and runtime device handling

Support optional ref_audio samples in finetuning and make runtime device selection explicit while keeping auto fallback behavior consistent. Also ignore the local app override file to avoid accidental commits.

Made-with: Cursor

tests/test_cli.py

@@ -58,6 +58,7 @@ def test_parser_defaults_to_voxcpm2():
     parser = cli._build_parser()
     args = parser.parse_args(["design", "--text", "hello", "--output", "out.wav"])
     assert args.hf_model_id == "openbmb/VoxCPM2"
+    assert args.device == "auto"
     assert args.no_optimize is False
 
 
@@ -85,6 +86,7 @@ def test_load_model_respects_no_optimize_for_local_model(monkeypatch):
 
     cli.load_model(args)
 
+    assert calls["kwargs"]["device"] == "auto"
     assert calls["kwargs"]["optimize"] is False
 
 
@@ -110,6 +112,7 @@ def test_load_model_defaults_optimize_for_hf(monkeypatch):
 
     cli.load_model(args)
 
+    assert calls["kwargs"]["device"] == "auto"
     assert calls["kwargs"]["optimize"] is True
 
 
@@ -136,9 +139,37 @@ def test_load_model_respects_no_optimize_for_hf(monkeypatch):
 
     cli.load_model(args)
 
+    assert calls["kwargs"]["device"] == "auto"
     assert calls["kwargs"]["optimize"] is False
 
 
+def test_load_model_passes_explicit_device_to_hf(monkeypatch):
+    calls = {}
+
+    class FakeVoxCPM:
+        @classmethod
+        def from_pretrained(cls, **kwargs):
+            calls["kwargs"] = kwargs
+            return DummyModel()
+
+    monkeypatch.setattr(cli, "VoxCPM", FakeVoxCPM)
+    args = cli._build_parser().parse_args(
+        [
+            "design",
+            "--text",
+            "hello",
+            "--output",
+            "out.wav",
+            "--device",
+            "mps",
+        ]
+    )
+
+    cli.load_model(args)
+
+    assert calls["kwargs"]["device"] == "mps"
+
+
 def test_design_subcommand_applies_control(monkeypatch, tmp_path):
     dummy_model = DummyModel()
     monkeypatch.setattr(cli, "load_model", lambda args: dummy_model)

tests/test_model_utils.py

@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+UTILS_PATH = ROOT / "src" / "voxcpm" / "model" / "utils.py"
+
+transformers_stub = types.ModuleType("transformers")
+transformers_stub.PreTrainedTokenizer = object
+sys.modules.setdefault("transformers", transformers_stub)
+
+spec = importlib.util.spec_from_file_location("voxcpm.model.utils", UTILS_PATH)
+utils = importlib.util.module_from_spec(spec)
+assert spec.loader is not None
+spec.loader.exec_module(utils)
+
+
+def test_resolve_runtime_device_auto_falls_back_to_cpu(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: False)
+    monkeypatch.setattr(utils, "_has_mps", lambda: False)
+
+    assert utils.resolve_runtime_device(None, "cuda") == "cpu"
+
+
+def test_resolve_runtime_device_auto_uses_mps_when_available(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: False)
+    monkeypatch.setattr(utils, "_has_mps", lambda: True)
+
+    assert utils.resolve_runtime_device("auto", "cuda") == "mps"
+
+
+def test_resolve_runtime_device_respects_explicit_cpu(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: True)
+    monkeypatch.setattr(utils, "_has_mps", lambda: True)
+
+    assert utils.resolve_runtime_device("cpu", "cuda") == "cpu"
+
+
+def test_resolve_runtime_device_rejects_unavailable_explicit_cuda(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: False)
+    monkeypatch.setattr(utils, "_has_mps", lambda: True)
+
+    with pytest.raises(ValueError, match="CUDA is not available"):
+        utils.resolve_runtime_device("cuda:0", "cuda")