OpenBMB/VoxCPM · error · ValueError

Requested device '{device}', but CUDA is not available. Use

Error message

Requested device '{device}', but CUDA is not available. Use device='auto' for automatic fallback.

What it means

resolve_runtime_device treats 'cuda'/'cuda:N' as an explicit requirement: if torch.cuda.is_available() is False it raises rather than silently falling back (use 'auto' for fallback).

Source

Thrown at src/voxcpm/model/utils.py:226

    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'."
    )

View on GitHub (pinned to f5a1c6a6b9)

Solutions

  1. Use device='auto' to let it pick the best available device
  2. Install/fix CUDA-enabled torch (pip install torch with cuda wheels) and drivers
  3. If CPU is intended, pass device='cpu' explicitly

Example fix

# before
model = VoxCPM(..., device="cuda")
# after
model = VoxCPM(..., device="auto")
Defensive patterns

Strategy: fallback

Validate before calling

import torch
if device and device.startswith("cuda") and not torch.cuda.is_available():
    device = "auto"

Type guard

def cuda_ok() -> bool:
    import torch; return torch.cuda.is_available()

Prevention

When it happens

Trigger: Passing device='cuda' on a machine without CUDA GPU, without NVIDIA drivers, or in a CPU-only Docker/CI image; also broken CUDA installs where torch cannot initialize it.

Common situations: Developing on Mac/CI then deploying config with device='cuda'; driver/CUDA toolkit mismatch; GPU occupied/unavailable in containers without nvidia runtime.

Related errors

AI-assisted analysis of OpenBMB/VoxCPM@f5a1c6a6b9 (2026-08-27). Data as JSON: /api/errors/49ae619a824b86dd. Report an issue: GitHub.