Flatten (TorchLean.Tensor → FloatArray)

Row-major order with outermost axis first and innermost axis last.

def Runtime.Autograd.Cuda.Convert.Internal.flattenFloat {s : Spec.Shape} : TorchLean.Tensor Float s → FloatArray → FloatArray

Append the elements of a tensor to an accumulator in row-major order.

def Runtime.Autograd.Cuda.Convert.flattenFloat {s : Spec.Shape} (t : TorchLean.Tensor Float s) : FloatArray

Flatten a TorchLean.Tensor Float s into a row-major FloatArray (CUDA-compatible).

def Runtime.Autograd.Cuda.Convert.Internal.flattenBoolMask {s : Spec.Shape} : TorchLean.Tensor Bool s → FloatArray → FloatArray

Accumulating worker for flattenBoolMask: true becomes 1.0, false becomes 0.0, since the CUDA side has no boolean buffers and reads masks as floats.

def Runtime.Autograd.Cuda.Convert.flattenBoolMask {s : Spec.Shape} (mask : TorchLean.Tensor Bool s) : FloatArray

Flatten a TorchLean.Tensor Bool s mask to FloatArray as 0.0/1.0 values in row-major order.

Unflatten (FloatArray → TorchLean.Tensor)

These functions assume row-major order. The public operations check the expected length and return none on mismatch. Runtime code that already owns the buffer-size invariant calls the internal workers directly.

def Runtime.Autograd.Cuda.Convert.Internal.unflattenFloat {s : Spec.Shape} (a : FloatArray) (offset : ℕ) : TorchLean.Tensor Float s

Read a tensor out of a starting at offset, trusting the caller for the bounds. The checked entry point is unflattenFloat?; this one exists so a batched download can slice one array.

def Runtime.Autograd.Cuda.Convert.unflattenFloat? {s : Spec.Shape} (a : FloatArray) : Option (TorchLean.Tensor Float s)

Unflatten a row-major FloatArray into a TorchLean.Tensor Float s when a.size matches.