opendatalab/MinerU · error · ValueError

CUDA is not available.

Error message

CUDA is not available.

What it means

Raised by set_lmdeploy_backend() in the VLM backend when device_type is 'cuda' but torch.cuda.is_available() is False. The function needs a working CUDA runtime to then choose turbomind vs pytorch backend by OS and compute capability, so a CUDA-less environment fails fast here.

Source

Thrown at mineru/backend/vlm/utils.py:66

    elif version.parse(compute_capability) < version.parse("8.0"):
        if version.parse(vllm_version) >= version.parse("0.10.2"):
            logger.info(f"compute_capability: {compute_capability} < 8.0, but vllm version: {vllm_version} >= 0.10.2, enable custom_logits_processors")
            return True
        else:
            logger.info(f"compute_capability: {compute_capability} < 8.0 and vllm version: {vllm_version} < 0.10.2, disable custom_logits_processors")
            return False
    else:
        logger.info(f"compute_capability: {compute_capability} >= 8.0 and vllm version: {vllm_version} >= 0.10.1, enable custom_logits_processors")
        return True


def set_lmdeploy_backend(device_type: str) -> str:
    if device_type.lower() in ["ascend", "maca", "camb"]:
        lmdeploy_backend = "pytorch"
    elif device_type.lower() in ["cuda"]:
        import torch
        if not torch.cuda.is_available():
            raise ValueError("CUDA is not available.")
        if is_windows_environment():
            lmdeploy_backend = "turbomind"
        elif is_linux_environment():
            major, minor = torch.cuda.get_device_capability()
            compute_capability = f"{major}.{minor}"
            if version.parse(compute_capability) >= version.parse("8.0"):
                lmdeploy_backend = "pytorch"
            else:
                lmdeploy_backend = "turbomind"
        else:
            raise ValueError("Unsupported operating system.")
    else:
        raise ValueError(f"Unsupported lmdeploy device type: {device_type}")
    return lmdeploy_backend


def set_default_gpu_memory_utilization() -> float:
    from vllm import __version__ as vllm_version

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Check torch.cuda.is_available() and torch.version.cuda in your environment.
2. If GPU exists: install a CUDA-enabled torch build matching your driver, and verify nvidia-smi works.
3. If no GPU: use a different backend (e.g. transformers on CPU/MPS, mlx-engine on Apple Silicon, or an http-client backend pointed at a GPU server).
4. Unset or fix CUDA_VISIBLE_DEVICES.

Example fix

# before
backend = set_lmdeploy_backend("cuda")  # ValueError: CUDA is not available.

# after
import torch
assert torch.cuda.is_available(), "need CUDA-enabled torch + driver"
backend = set_lmdeploy_backend("cuda")

import torch

def cuda_ready() -> bool:
    return torch.cuda.is_available()

if not cuda_ready():
    choose_non_lmdeploy_backend()  # transformers / http-client

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

try:
    backend = set_lmdeploy_backend("cuda")
except ValueError as e:
    if "CUDA is not available" in str(e):
        raise RuntimeError("install CUDA-enabled torch or pick another backend") from e
    raise

When it happens

Trigger: device_type='cuda' passed to (or defaulted for) lmdeploy engine setup on a machine with no NVIDIA GPU, nvidia drivers, or a CPU-only torch build; CUDA_VISIBLE_DEVICES set to an empty/invalid value.

Common situations: CPU-only torch wheel installed in a GPU container; driver/library mismatch so torch cannot initialize CUDA; remote device hidden by CUDA_VISIBLE_DEVICES="".

Related errors

• Unsupported lmdeploy device type: {device_type}
• Unsupported lmdeploy device type: {device_type}
• NPU is selected as device, but torch_npu is not available. P
• NPU is selected as device, but torch_npu is not available. P
• Unsupported operating system.

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/a28c414199a6d4bd. Report an issue: GitHub.