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

.gitignore

@@ -3,3 +3,4 @@ __pycache__
 voxcpm.egg-info
 .DS_Store
 ./pretrained_models/
+app_local.py

src/voxcpm/cli.py

@@ -209,6 +209,7 @@ def load_model(args) -> VoxCPM:
                 zipenhancer_model_path=zipenhancer_path,
                 enable_denoiser=not args.no_denoiser,
                 optimize=not args.no_optimize,
+                device=args.device,
                 lora_config=lora_config,
                 lora_weights_path=lora_weights_path,
             )
@@ -227,6 +228,7 @@ def load_model(args) -> VoxCPM:
             cache_dir=args.cache_dir,
             local_files_only=args.local_files_only,
             optimize=not args.no_optimize,
+            device=args.device,
             lora_config=lora_config,
             lora_weights_path=lora_weights_path,
         )
@@ -403,6 +405,12 @@ def _add_model_args(parser):
         default=DEFAULT_HF_MODEL_ID,
         help=f"Hugging Face repo id (default: {DEFAULT_HF_MODEL_ID})",
     )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="auto",
+        help="Runtime device: auto, cpu, mps, cuda, or cuda:N (default: auto)",
+    )
     parser.add_argument(
         "--cache-dir", type=str, help="Cache directory for Hub downloads"
     )

src/voxcpm/core.py

@@ -17,6 +17,7 @@ class VoxCPM:
         zipenhancer_model_path: str | None = "iic/speech_zipenhancer_ans_multiloss_16k_base",
         enable_denoiser: bool = True,
         optimize: bool = True,
+        device: str | None = None,
         lora_config: Optional[LoRAConfig] = None,
         lora_weights_path: Optional[str] = None,
     ):
@@ -30,6 +31,9 @@ class VoxCPM:
                 id or local path. If None, denoiser will not be initialized.
             enable_denoiser: Whether to initialize the denoiser pipeline.
             optimize: Whether to optimize the model with torch.compile. True by default, but can be disabled for debugging.
+            device: Runtime device. If set to ``None`` or ``"auto"``, VoxCPM
+                will choose automatically (preferring CUDA, then MPS, then CPU).
+                If set explicitly, that device is used or a clear error is raised.
             lora_config: LoRA configuration for fine-tuning. If lora_weights_path is
                 provided without lora_config, a default config will be created.
             lora_weights_path: Path to pre-trained LoRA weights (.pth file or directory
@@ -56,10 +60,20 @@ class VoxCPM:
         arch = config.get("architecture", "voxcpm").lower()
 
         if arch == "voxcpm2":
-            self.tts_model = VoxCPM2Model.from_local(voxcpm_model_path, optimize=optimize, lora_config=lora_config)
+            self.tts_model = VoxCPM2Model.from_local(
+                voxcpm_model_path,
+                optimize=optimize,
+                device=device,
+                lora_config=lora_config,
+            )
             print("Loaded VoxCPM2Model", file=sys.stderr)
         elif arch == "voxcpm":
-            self.tts_model = VoxCPMModel.from_local(voxcpm_model_path, optimize=optimize, lora_config=lora_config)
+            self.tts_model = VoxCPMModel.from_local(
+                voxcpm_model_path,
+                optimize=optimize,
+                device=device,
+                lora_config=lora_config,
+            )
             print("Loaded VoxCPMModel", file=sys.stderr)
         else:
             raise ValueError(f"Unsupported architecture: {arch}")
@@ -94,6 +108,7 @@ class VoxCPM:
         cache_dir: str = None,
         local_files_only: bool = False,
         optimize: bool = True,
+        device: str | None = None,
         lora_config: Optional[LoRAConfig] = None,
         lora_weights_path: Optional[str] = None,
         **kwargs,
@@ -109,6 +124,9 @@ class VoxCPM:
             cache_dir: Custom cache directory for the snapshot.
             local_files_only: If True, only use local files and do not attempt
                 to download.
+            device: Runtime device. Use ``None``/``"auto"`` for automatic
+                fallback, or an explicit value such as ``"cpu"``, ``"mps"``,
+                ``"cuda"``, or ``"cuda:0"``.
             lora_config: LoRA configuration for fine-tuning. If lora_weights_path is
                 provided without lora_config, a default config will be created with
                 enable_lm=True and enable_dit=True.
@@ -146,6 +164,7 @@ class VoxCPM:
             zipenhancer_model_path=zipenhancer_model_id if load_denoiser else None,
             enable_denoiser=load_denoiser,
             optimize=optimize,
+            device=device,
             lora_config=lora_config,
             lora_weights_path=lora_weights_path,
             **kwargs,

src/voxcpm/model/utils.py

@@ -1,4 +1,4 @@
-from typing import List
+from typing import List, Optional
 import torch
 from transformers import PreTrainedTokenizer
 
@@ -119,3 +119,67 @@ def get_dtype(dtype: str):
         return torch.float32
     else:
         raise ValueError(f"Unsupported dtype: {dtype}")
+
+
+def _has_mps() -> bool:
+    return hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+
+
+def auto_select_device(preferred_device: Optional[str] = "cuda") -> str:
+    """
+    Choose a runtime device automatically.
+
+    Preference order:
+    - if the preferred device is available, use it
+    - otherwise fall back to CUDA -> MPS -> CPU
+    """
+    preferred = (preferred_device or "cuda").strip().lower()
+
+    if preferred.startswith("cuda") and torch.cuda.is_available():
+        return preferred
+    if preferred == "mps" and _has_mps():
+        return "mps"
+    if preferred == "cpu":
+        return "cpu"
+
+    if torch.cuda.is_available():
+        return "cuda"
+    if _has_mps():
+        return "mps"
+    return "cpu"
+
+
+def resolve_runtime_device(device: Optional[str], configured_device: str = "cuda") -> str:
+    """
+    Resolve the actual runtime device.
+
+    Semantics:
+    - ``device`` is ``None`` or ``"auto"``: use automatic fallback selection
+    - otherwise: treat it as an explicit user choice and validate availability
+    """
+    explicit = None if device is None else device.strip().lower()
+
+    if explicit is None or explicit == "auto":
+        return auto_select_device(configured_device)
+
+    if explicit.startswith("cuda"):
+        if not torch.cuda.is_available():
+            raise ValueError(
+                f"Requested device '{device}', but CUDA is not available. "
+                "Use device='auto' for automatic fallback."
+            )
+        return explicit
+    if explicit == "mps":
+        if not _has_mps():
+            raise ValueError(
+                "Requested device 'mps', but MPS is not available. "
+                "Use device='auto' for automatic fallback."
+            )
+        return "mps"
+    if explicit == "cpu":
+        return "cpu"
+
+    raise ValueError(
+        f"Unsupported device '{device}'. Supported values are 'auto', 'cpu', 'mps', "
+        "'cuda', or indexed CUDA devices like 'cuda:0'."
+    )

src/voxcpm/model/voxcpm.py

@@ -44,7 +44,7 @@ from ..modules.layers.lora import apply_lora_to_named_linear_modules
 from ..modules.locdit import CfmConfig, UnifiedCFM, VoxCPMLocDiT
 from ..modules.locenc import VoxCPMLocEnc
 from ..modules.minicpm4 import MiniCPM4Config, MiniCPMModel
-from .utils import get_dtype, mask_multichar_chinese_tokens
+from .utils import get_dtype, mask_multichar_chinese_tokens, resolve_runtime_device
 
 
 class VoxCPMEncoderConfig(BaseModel):
@@ -109,18 +109,15 @@ class VoxCPMModel(nn.Module):
         tokenizer: LlamaTokenizerFast,
         audio_vae: AudioVAE,
         lora_config: LoRAConfig = None,
+        device: str | None = None,
     ):
         super().__init__()
         self.config = config
         self.lora_config = lora_config
         self.feat_dim = config.feat_dim
         self.patch_size = config.patch_size
-        self.device = config.device
+        self.device = resolve_runtime_device(device, config.device)
-        if not torch.cuda.is_available():
+        self.config.device = self.device
-            if torch.backends.mps.is_available():
-                self.device = "mps"
-            else:
-                self.device = "cpu"
         print(f"Running on device: {self.device}, dtype: {self.config.dtype}", file=sys.stderr)
 
         # Text-Semantic LM
@@ -847,7 +844,14 @@ class VoxCPMModel(nn.Module):
             yield feat_pred, pred_feat_seq.squeeze(0).cpu()
 
     @classmethod
-    def from_local(cls, path: str, optimize: bool = True, training: bool = False, lora_config: LoRAConfig = None):
+    def from_local(
+        cls,
+        path: str,
+        optimize: bool = True,
+        training: bool = False,
+        device: str | None = None,
+        lora_config: LoRAConfig = None,
+    ):
         config = VoxCPMConfig.model_validate_json(open(os.path.join(path, "config.json")).read())
         tokenizer = LlamaTokenizerFast.from_pretrained(path)
         audio_vae_config = getattr(config, "audio_vae_config", None)
@@ -870,7 +874,7 @@ class VoxCPMModel(nn.Module):
             raise FileNotFoundError(
                 f"AudioVAE checkpoint not found. Expected either {audiovae_safetensors_path} or {audiovae_pth_path}"
             )
-        model = cls(config, tokenizer, audio_vae, lora_config)
+        model = cls(config, tokenizer, audio_vae, lora_config, device=device)
         if not training:
             lm_dtype = get_dtype(model.config.dtype)
             model = model.to(lm_dtype)

src/voxcpm/model/voxcpm2.py

@@ -45,7 +45,7 @@ from ..modules.layers.lora import apply_lora_to_named_linear_modules
 from ..modules.locdit import CfmConfig, UnifiedCFM, VoxCPMLocDiTV2
 from ..modules.locenc import VoxCPMLocEnc
 from ..modules.minicpm4 import MiniCPM4Config, MiniCPMModel
-from .utils import get_dtype, mask_multichar_chinese_tokens
+from .utils import get_dtype, mask_multichar_chinese_tokens, resolve_runtime_device
 
 
 # A simple function to trim audio silence using VAD, not used default
@@ -151,18 +151,15 @@ class VoxCPM2Model(nn.Module):
         tokenizer: LlamaTokenizerFast,
         audio_vae: AudioVAEV2,
         lora_config: LoRAConfig = None,
+        device: str | None = None,
     ):
         super().__init__()
         self.config = config
         self.lora_config = lora_config
         self.feat_dim = config.feat_dim
         self.patch_size = config.patch_size
-        self.device = config.device
+        self.device = resolve_runtime_device(device, config.device)
-        if not torch.cuda.is_available():
+        self.config.device = self.device
-            if torch.backends.mps.is_available():
-                self.device = "mps"
-            else:
-                self.device = "cpu"
         print(f"Running on device: {self.device}, dtype: {self.config.dtype}", file=sys.stderr)
 
         # Text-Semantic LM
@@ -1098,7 +1095,14 @@ class VoxCPM2Model(nn.Module):
             yield feat_pred, generated_feat, context_len
 
     @classmethod
-    def from_local(cls, path: str, optimize: bool = True, training: bool = False, lora_config: LoRAConfig = None):
+    def from_local(
+        cls,
+        path: str,
+        optimize: bool = True,
+        training: bool = False,
+        device: str | None = None,
+        lora_config: LoRAConfig = None,
+    ):
         config = VoxCPMConfig.model_validate_json(open(os.path.join(path, "config.json")).read())
         tokenizer = LlamaTokenizerFast.from_pretrained(path)
         audio_vae_config = getattr(config, "audio_vae_config", None)
@@ -1121,7 +1125,7 @@ class VoxCPM2Model(nn.Module):
             raise FileNotFoundError(
                 f"AudioVAE checkpoint not found. Expected either {audiovae_safetensors_path} or {audiovae_pth_path}"
             )
-        model = cls(config, tokenizer, audio_vae, lora_config)
+        model = cls(config, tokenizer, audio_vae, lora_config, device=device)
         if not training:
             lm_dtype = get_dtype(model.config.dtype)
             model = model.to(lm_dtype)

src/voxcpm/training/data.py

@@ -12,6 +12,7 @@ from .packers import AudioFeatureProcessingPacker
 
 DEFAULT_TEXT_COLUMN = "text"
 DEFAULT_AUDIO_COLUMN = "audio"
+DEFAULT_REF_AUDIO_COLUMN = "ref_audio"
 DEFAULT_ID_COLUMN = "dataset_id"
 
 
@@ -21,6 +22,7 @@ def load_audio_text_datasets(
     val_manifest: str = "",
     text_column: str = DEFAULT_TEXT_COLUMN,
     audio_column: str = DEFAULT_AUDIO_COLUMN,
+    ref_audio_column: str = DEFAULT_REF_AUDIO_COLUMN,
     dataset_id_column: str = DEFAULT_ID_COLUMN,
     sample_rate: int = 16_000,
     num_proc: int = 1,
@@ -34,14 +36,19 @@ def load_audio_text_datasets(
     def prepare(ds: Dataset) -> Dataset:
         if audio_column not in ds.column_names:
             raise ValueError(f"Expected '{audio_column}' column in manifest.")
-        # We cast to Audio to ensure proper handling during training,
-        # but for length calculation we might need raw path or duration if available.
-        # HF datasets usually don't compute duration automatically for 'Audio' column.
         ds = ds.cast_column(audio_column, Audio(sampling_rate=sample_rate))
         if audio_column != DEFAULT_AUDIO_COLUMN:
             ds = ds.rename_column(audio_column, DEFAULT_AUDIO_COLUMN)
         if text_column != DEFAULT_TEXT_COLUMN:
             ds = ds.rename_column(text_column, DEFAULT_TEXT_COLUMN)
+
+        # ref_audio is optional — cast to Audio if the column exists
+        ref_col = ref_audio_column if ref_audio_column in ds.column_names else DEFAULT_REF_AUDIO_COLUMN
+        if ref_col in ds.column_names:
+            ds = ds.cast_column(ref_col, Audio(sampling_rate=sample_rate))
+            if ref_col != DEFAULT_REF_AUDIO_COLUMN:
+                ds = ds.rename_column(ref_col, DEFAULT_REF_AUDIO_COLUMN)
+
         if dataset_id_column and dataset_id_column in ds.column_names:
             if dataset_id_column != DEFAULT_ID_COLUMN:
                 ds = ds.rename_column(dataset_id_column, DEFAULT_ID_COLUMN)
@@ -67,11 +74,11 @@ def compute_sample_lengths(
     - 音频长度:
         duration(s) * audio_vae_fps -> 近似 VAE 帧数 t_vae
         t_seq = ceil(t_vae / patch_size)
-    - 序列总长约为: text_len + t_seq + 2
+    - 无 ref_audio: text_len + t_seq + 2
+    - 有 ref_audio: text_len + t_seq + ref_seq + 4
 
     Optimized: Use batch column access instead of iterating item by item.
     """
-    # Batch access columns - much faster than per-item access
     text_ids_list = ds["text_ids"]
     text_lens = [len(t) for t in text_ids_list]
 
@@ -79,18 +86,35 @@ def compute_sample_lengths(
     if has_duration:
         durations = ds["duration"]
     else:
-        # Fallback: need to compute from audio (slow, but unavoidable without duration column)
         durations = []
         for i in range(len(ds)):
             audio = ds[i][DEFAULT_AUDIO_COLUMN]
             durations.append(len(audio["array"]) / float(audio["sampling_rate"]))
 
-    # Vectorized length computation
+    has_ref_audio = DEFAULT_REF_AUDIO_COLUMN in ds.column_names
+    if has_ref_audio:
+        ref_duration_col = "ref_duration" if "ref_duration" in ds.column_names else None
+
     lengths = []
-    for text_len, duration in zip(text_lens, durations):
+    for i, (text_len, duration) in enumerate(zip(text_lens, durations)):
         t_vae = math.ceil(float(duration) * audio_vae_fps)
         t_seq = math.ceil(t_vae / patch_size)
-        total_len = text_len + t_seq + 2
+
+        ref_seq = 0
+        if has_ref_audio:
+            # Estimate ref_audio length; ref_audio is None for samples without it
+            if ref_duration_col:
+                ref_dur = ds[i].get(ref_duration_col)
+            else:
+                ref_item = ds[i].get(DEFAULT_REF_AUDIO_COLUMN)
+                ref_dur = len(ref_item["array"]) / float(ref_item["sampling_rate"]) if ref_item else None
+            if ref_dur is not None and float(ref_dur) > 0:
+                ref_vae = math.ceil(float(ref_dur) * audio_vae_fps)
+                ref_seq = math.ceil(ref_vae / patch_size)
+
+        # +2 for 101/102; +2 more for 103/104 when ref_audio present
+        overhead = 4 if ref_seq > 0 else 2
+        total_len = text_len + t_seq + ref_seq + overhead
         lengths.append(total_len)
 
     return lengths
@@ -102,8 +126,11 @@ class HFVoxCPMDataset(TorchDataset):
     PyTorch-friendly samples.
     """
 
+    _SENTINEL = [-100.0]
+
     def __init__(self, dataset: Dataset):
         self.dataset = dataset
+        self.has_ref_audio = DEFAULT_REF_AUDIO_COLUMN in dataset.column_names
 
     def __len__(self):
         return len(self.dataset)
@@ -111,13 +138,17 @@ class HFVoxCPMDataset(TorchDataset):
     def __getitem__(self, idx: int):
         item = self.dataset[idx]
         audio = item[DEFAULT_AUDIO_COLUMN]
-        return {
+        sample = {
             "text_ids": item["text_ids"],
             "audio_array": audio["array"],
             "audio_sampling_rate": audio["sampling_rate"],
             "dataset_id": item.get(DEFAULT_ID_COLUMN, 0),
             "is_prompt": item.get("is_prompt", False),
         }
+        if self.has_ref_audio:
+            ref = item.get(DEFAULT_REF_AUDIO_COLUMN)
+            sample["ref_audio_array"] = ref["array"] if ref else self._SENTINEL
+        return sample
 
     @staticmethod
     def pad_sequences(seqs: List[torch.Tensor], pad_value: float):
@@ -143,7 +174,7 @@ class HFVoxCPMDataset(TorchDataset):
         audio_padded = cls.pad_sequences(audio_tensors, pad_value=-100.0)
         task_ids = torch.ones(text_padded.size(0), dtype=torch.int32)
 
-        return {
+        result = {
             "text_tokens": text_padded,
             "audio_tokens": audio_padded,
             "task_ids": task_ids,
@@ -151,6 +182,12 @@ class HFVoxCPMDataset(TorchDataset):
             "is_prompts": is_prompts,
         }
 
+        if "ref_audio_array" in batch[0]:
+            ref_tensors = [torch.tensor(s["ref_audio_array"], dtype=torch.float32) for s in batch]
+            result["ref_audio_tokens"] = cls.pad_sequences(ref_tensors, pad_value=-100.0)
+
+        return result
+
 
 class BatchProcessor:
     """
@@ -184,12 +221,17 @@ class BatchProcessor:
         task_ids = batch["task_ids"].to(self.device)
         dataset_ids = batch["dataset_ids"].to(self.device)
 
+        ref_audio_tokens = None
+        if "ref_audio_tokens" in batch:
+            ref_audio_tokens = batch["ref_audio_tokens"].to(self.device)
+
         packed = self.packer(
             audio_tokens=audio_tokens,
             text_tokens=text_tokens,
             task_ids=task_ids,
             dataset_ids=dataset_ids,
             is_prompts=batch["is_prompts"],
+            ref_audio_tokens=ref_audio_tokens,
         )
         return packed
 

src/voxcpm/training/packers.py

@@ -1,4 +1,4 @@
-from typing import Dict, List
+from typing import Dict, List, Optional
 
 import torch
 import torch.nn as nn
@@ -14,7 +14,6 @@ class AudioFeatureProcessingPacker:
     def __init__(self, dataset_cnt: int, max_len: int, patch_size: int, feat_dim: int, audio_vae: nn.Module):
         self.audio_start_id = 101
         self.audio_end_id = 102
-        # unused now
         self.audio_prompt_start_id = 103
         self.audio_prompt_end_id = 104
         self.text_eos_token_id = 2
@@ -78,11 +77,16 @@ class AudioFeatureProcessingPacker:
         task_ids: torch.Tensor,
         dataset_ids: torch.Tensor,
         is_prompts: List[bool],
+        ref_audio_tokens: Optional[torch.Tensor] = None,
     ) -> Dict[str, torch.Tensor]:
         """
         Padding-based batching: each sample in the input batch is processed
         independently and then padded to a common length (capped by ``max_len``).
         The result tensors all have shape [B, T, ...].
+
+        If ``ref_audio_tokens`` is provided (same batch dim as ``audio_tokens``),
+        samples whose unpadded ref_audio length > 0 will be processed with the
+        reference-audio path (tokens 103/104 prepended, loss only on target audio).
         """
         device = audio_tokens.device
         max_dataset_id = int(dataset_ids.max().item()) if dataset_ids.numel() > 0 else -1
@@ -101,23 +105,43 @@ class AudioFeatureProcessingPacker:
         audio_duration_consumed = torch.zeros(dataset_cnt, dtype=torch.float32, device=device)
         text_token_consumed = torch.zeros(dataset_cnt, dtype=torch.float32, device=device)
 
-        for audio_token, text_token, task_id, dataset_idx, is_prompt in zip(
+        ref_iter = ref_audio_tokens if ref_audio_tokens is not None else [None] * audio_tokens.size(0)
-            audio_tokens, text_tokens, task_ids.tolist(), dataset_ids.tolist(), is_prompts
+
+        for audio_token, text_token, task_id, dataset_idx, is_prompt, ref_token in zip(
+            audio_tokens, text_tokens, task_ids.tolist(), dataset_ids.tolist(), is_prompts, ref_iter
         ):
             unpad_audio_token = self.unpad_audio_tokens(audio_token).to(torch.float32)
             unpad_text_token = self.unpad_text_tokens(text_token)
             usage = self.id_to_task[task_id]
 
-            (
+            has_ref = False
-                packed_text,
+            if ref_token is not None:
-                audio_feat,
+                unpad_ref_token = self.unpad_audio_tokens(ref_token).to(torch.float32)
-                text_mask,
+                if unpad_ref_token.numel() > 0:
-                audio_mask,
+                    has_ref = True
-                loss_mask,
+
-                labels,
+            if has_ref:
-                audio_duration,
+                (
-                text_token_count,
+                    packed_text,
-            ) = self.process_functions[usage](unpad_audio_token, unpad_text_token, is_prompt)
+                    audio_feat,
+                    text_mask,
+                    audio_mask,
+                    loss_mask,
+                    labels,
+                    audio_duration,
+                    text_token_count,
+                ) = self.process_tts_data_with_ref(unpad_ref_token, unpad_audio_token, unpad_text_token)
+            else:
+                (
+                    packed_text,
+                    audio_feat,
+                    text_mask,
+                    audio_mask,
+                    loss_mask,
+                    labels,
+                    audio_duration,
+                    text_token_count,
+                ) = self.process_functions[usage](unpad_audio_token, unpad_text_token, is_prompt)
 
             audio_duration_consumed[dataset_idx] += audio_duration
             text_token_consumed[dataset_idx] += text_token_count
@@ -294,3 +318,89 @@ class AudioFeatureProcessingPacker:
             audio_duration,
             text_token_count,
         )
+
+    def process_tts_data_with_ref(
+        self,
+        ref_audio_token: torch.Tensor,
+        target_audio_token: torch.Tensor,
+        text_token: torch.Tensor,
+    ):
+        """
+        Build a training sequence with reference audio prepended:
+
+            [103, ref_feats, 104, text, 101, target_feats, 102]
+
+        Loss is computed only on the target audio segment.
+        """
+        device = text_token.device
+        txt_len = len(text_token)
+
+        ref_feats, ref_duration = self.extract_audio_feats(ref_audio_token)
+        ref_feats = ref_feats.squeeze(0)  # [R, P, D]
+        ref_len = ref_feats.shape[0]
+
+        tgt_feats, tgt_duration = self.extract_audio_feats(target_audio_token)
+        tgt_feats = tgt_feats.squeeze(0)  # [A, P, D]
+        tgt_len = tgt_feats.shape[0]
+
+        feat_shape = (self.patch_size, ref_feats.size(-1))
+
+        def _tok(ids):
+            return torch.tensor(ids, dtype=torch.int32, device=device)
+
+        # -- text token track --
+        # [103, 0×R, 104, text_ids, 101, 0×A, 102]
+        text_token_info = torch.cat([
+            _tok([self.audio_prompt_start_id]),
+            torch.zeros(ref_len, dtype=torch.int32, device=device),
+            _tok([self.audio_prompt_end_id]),
+            text_token,
+            _tok([self.audio_start_id]),
+            torch.zeros(tgt_len, dtype=torch.int32, device=device),
+            _tok([self.audio_end_id]),
+        ])
+
+        # -- audio feature track --
+        zero_1 = torch.zeros((1,) + feat_shape, dtype=torch.float32, device=device)
+        zero_txt = torch.zeros((txt_len,) + feat_shape, dtype=torch.float32, device=device)
+        audio_feat_info = torch.cat([
+            zero_1, ref_feats, zero_1,      # 103, ref, 104
+            zero_txt,                        # text
+            zero_1, tgt_feats, zero_1,       # 101, target, 102
+        ], dim=0)
+
+        # -- masks --
+        text_mask = torch.cat([
+            torch.ones(1), torch.zeros(ref_len), torch.ones(1),
+            torch.ones(txt_len),
+            torch.ones(1), torch.zeros(tgt_len), torch.ones(1),
+        ]).to(torch.int32).to(device)
+
+        audio_mask = torch.cat([
+            torch.zeros(1), torch.ones(ref_len), torch.zeros(1),
+            torch.zeros(txt_len),
+            torch.zeros(1), torch.ones(tgt_len), torch.zeros(1),
+        ]).to(torch.int32).to(device)
+
+        loss_mask = torch.cat([
+            torch.zeros(1 + ref_len + 1),   # ref part: no loss
+            torch.zeros(txt_len),            # text: no loss
+            torch.zeros(1),                  # 101: no loss
+            torch.ones(tgt_len),             # target audio: LOSS
+            torch.zeros(1),                  # 102: no loss
+        ]).to(torch.int32).to(device)
+
+        total_len = 1 + ref_len + 1 + txt_len + 1 + tgt_len + 1
+        labels = torch.zeros(total_len, dtype=torch.int32, device=device)
+        labels[-2] = 1  # stop label at last target audio position
+
+        return (
+            text_token_info,
+            audio_feat_info,
+            text_mask,
+            audio_mask,
+            loss_mask,
+            labels,
+            ref_duration + tgt_duration,
+            txt_len,
+        )

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")