torch.js has not been released yet.

torch.WebGPUTensorData

export interface WebGPUTensorData {
  buffer: GPUBuffer;
  shape: readonly number[];
  dtype: DType;
  device: 'webgpu';
  requires_grad: boolean;
  grad_fn?: GradFn;
  /** If false, this tensor is a view and does not own its buffer (won't release on destroy) */
  ownsBuffer?: boolean;
}

buffer(GPUBuffer)
shape(readonly number[])
dtype(DType)
device('webgpu')
requires_grad(boolean)
grad_fn(GradFn)optional
ownsBuffer(boolean)optional: – If false, this tensor is a view and does not own its buffer (won't release on destroy)

Internal tensor data structure for WebGPU tensors.

Contains the metadata and GPU buffer reference for tensors allocated on the WebGPU device. This is the low-level representation that backs the public Tensor class.

Fields:

buffer: GPU buffer handle (GPUBuffer from WebGPU API)
shape: Dimensions of the tensor
dtype: Element data type
device: Always 'webgpu' for this variant
requires_grad: Whether gradients are tracked for this tensor
grad_fn: Function that computes gradients during backward pass
ownsBuffer: Whether this tensor owns the buffer (vs borrowed/view)

WebGPUOnlyResult