Crate grafos_tensor 

grafos-tensor – Tensor operations backed by leased fabric memory.

This crate provides FabricTensor, an N-dimensional tensor whose data is stored as contiguous row-major f32 values backed by fabric memory acquired through FBMU (Fabric Bootstrap Memory Unit) leases. Every constructor validates that the fabric has sufficient capacity before allocating, and holds the lease for the lifetime of the tensor.

Supported operations

Operation	Method	Constraints
Matrix multiply	FabricTensor::matmul	Both 2-D; inner dims match
Elementwise add	FabricTensor::add	Same shape
Elementwise mul	FabricTensor::mul	Same shape
Scalar multiply	FabricTensor::scale	Any shape
ReLU	FabricTensor::relu	Any shape
Softmax	FabricTensor::softmax	axis < ndim
Subtract	FabricTensor::subtract	Same shape
Sum axis	FabricTensor::sum_axis	axis < ndim
Sigmoid	FabricTensor::sigmoid	Any shape
Natural log	FabricTensor::ln	Any shape
Clamp	FabricTensor::clip	Any shape
Transpose	FabricTensor::transpose	ndim >= 2
Reshape	FabricTensor::reshape	Same total elements

Placement helpers:

• FabricTensor::to_gpu / FabricTensor::to_cpu
• FabricTensor::device, FabricTensor::is_cpu, FabricTensor::is_gpu

With gpu feature enabled, operations on GPU-placed tensors dispatch through the v1 fabricbios_gpu_v1 session surface — specifically a private submit_signal_kernel helper that opens a transient grafos_std::gpu::GpuSession on the tensor’s held GpuLease, module_loads the kernel binary, launches with [1,1,1] grid/block dims, syncs, and lets the session/module drop (RAII). Numerical outputs are still computed on the CPU path; GPU dispatch is control-path signaling only under the current mock kernels. See docs/grafos-tensor-guide.md for the full programming model.

Operator overloading is provided for &FabricTensor: + (elementwise add), * (elementwise mul), and * f32 (scalar mul).

Quick start

use grafos_tensor::FabricTensor;

let a = FabricTensor::from_slice(&[2, 3], &[1.0, 2.0, 3.0, 4.0, 5.0, 6.0]).unwrap();
let b = FabricTensor::from_slice(&[3, 2], &[7.0, 8.0, 9.0, 10.0, 11.0, 12.0]).unwrap();
let c = a.matmul(&b).unwrap();
assert_eq!(c.shape(), &[2, 2]);
assert_eq!(c.get(&[0, 0]).unwrap(), 58.0);

Operator overloading

use grafos_tensor::FabricTensor;

let a = FabricTensor::from_slice(&[3], &[1.0, 2.0, 3.0]).unwrap();
let b = FabricTensor::from_slice(&[3], &[4.0, 5.0, 6.0]).unwrap();

let sum = (&a + &b).unwrap();        // elementwise add
let prod = (&a * &b).unwrap();       // elementwise mul
let scaled = (&a * 2.0f32).unwrap(); // scalar mul

Testing

On native targets, initialize the mock FBMU backend before creating tensors:

grafos_std::host::reset_mock();
grafos_std::host::mock_set_fbmu_arena_size(1 << 20); // 1 MiB

Structs

FabricTensor

N-dimensional tensor backed by a fabric memory lease.

Shape

Shape metadata for an N-dimensional tensor.

Enums

Device

Placement of tensor data in the fabric.

Type Aliases

Result

Result alias using FabricError.