sgl-project/sglang · error · ValueError

MlxTensorView requires a Torch MPS tensor, got {owner.device

Error message

MlxTensorView requires a Torch MPS tensor, got {owner.device}

What it means

MlxTensorView is a lifetime-bound zero-copy view and only exists for MPS-backed torch tensors. Its constructor detaches the tensor and requires owner.device.type == 'mps', otherwise raising with the offending device.

Source

Thrown at python/sglang/srt/utils/tensor_bridge.py:118

class MlxTensorView:
    """A lifetime-bound, zero-copy MLX view of a Torch MPS tensor.

    The view deliberately retains a detached Torch tensor *and* the imported
    MLX array.  Holding only the array is insufficient: a later parameter
    replacement or garbage collection could invalidate the borrowed storage
    while MLX still has a lazy graph referring to it. This class is intended
    for immutable inference weights; construct a new view after replacing the
    source storage.
    """

    __slots__ = ("torch_tensor", "array")

    def __init__(self, tensor: torch.Tensor, *, synchronize: bool = True):
        with _BRIDGE_LOCK:
            owner = tensor.detach()
            if owner.device.type != "mps":
                raise ValueError(
                    f"MlxTensorView requires a Torch MPS tensor, got {owner.device}"
                )
            if synchronize:
                torch.mps.synchronize()
            self.torch_tensor = owner
            self.array = _torch_to_mlx(owner, copy=False, synchronize=False)

    @classmethod
    def _from_synchronized(cls, tensor: torch.Tensor) -> MlxTensorView:
        view = object.__new__(cls)
        owner = tensor.detach()
        if owner.device.type != "mps":
            raise ValueError(
                f"MlxTensorView requires a Torch MPS tensor, got {owner.device}"
            )
        view.torch_tensor = owner
        view.array = _torch_to_mlx(owner, copy=False, synchronize=False)
        return view

View on GitHub (pinned to 0132848349)

Solutions

1. Move the tensor to MPS first: MlxTensorView(t.to('mps'))
2. Use torch_to_mlx for CPU tensors (it copies) instead of a view

Example fix

# before
view = MlxTensorView(t)  # t on cpu
# after
view = MlxTensorView(t.to('mps'))

assert tensor.device.type == "mps", "MlxTensorView needs an MPS tensor"

def is_mps(t: torch.Tensor) -> bool:
    return t.device.type == "mps"

When it happens

Trigger: Constructing MlxTensorView(cpu_tensor) or MlxTensorView(cuda_tensor) directly.

Common situations: Unit tests or helper code constructing views on CPU fixtures; forgetting .to('mps') before wrapping.

Related errors

• SGLANG_USE_MLX requires stable Torch 2.13.x and MLX >= 0.32.
• The MLX tensor bridge supports CPU and MPS tensors, got {ten
• borrow_torch_tensors requires MPS tensors, got {devices}
• SGLANG_USE_MLX requires stable Torch 2.13.x and MLX >= 0.32.
• SGLANG_USE_MLX requires an available PyTorch MPS device

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2e1d37be170dc741. Report an issue: GitHub.