Module tensor 

Tensor — immutable, chainable wrapper around a libtorch tensor.

Every tensor owns its C++ handle and frees it on drop. This is the entire VRAM management story — no GC, no scopes, no finalizers.

Operations are chainable and return Result<Tensor>:

let z = a.add(&b)?.relu()?.sum()?;

Structs

DeviceInfo

Information about a CUDA device.

GradAccumulatorHandle

Opaque strong-reference handle to a leaf tensor’s AccumulateGrad node. Dropping it frees the node (unless a backward pass still holds its own reference).

RnnParams

Persistent cache of RNN parameter tensors on the C++ side.

Tensor

A tensor wrapping a libtorch C++ tensor.

TensorError

Error type for tensor operations.

TensorOptions

Options for tensor creation.

Enums

DType

Element data type of a tensor. Maps to PyTorch’s torch.dtype.

Device

Device represents where a tensor’s data lives.

Functions

cuda_active_bytes

Query bytes actively used by tensors on device 0.

cuda_active_bytes_idx

Query bytes actively used by tensors on a specific device.

cuda_allocated_bytes

Query bytes reserved by the CUDA caching allocator on device 0.

cuda_allocated_bytes_idx

Query bytes reserved by the CUDA caching allocator on a specific device.

cuda_available

Returns true if CUDA is available.

cuda_compute_capability

Query compute capability (major, minor) for a CUDA device.

cuda_device_count

Returns the number of CUDA devices.

cuda_device_name

Returns the GPU device name for device 0 (e.g. “NVIDIA GeForce GTX 1060 6GB”).

cuda_device_name_idx

Returns the GPU device name for the given index (e.g. “NVIDIA GeForce GTX 1060 6GB”).

cuda_devices

Enumerate all available CUDA devices.

cuda_empty_cache

Release all unused cached memory from the CUDA caching allocator. Equivalent to torch.cuda.empty_cache().

cuda_manual_seed_all

Seed all CUDA device RNGs. No-op when built without CUDA.

cuda_memory_info

Query CUDA memory usage for device 0. Returns (used_bytes, total_bytes) or an error if CUDA is not available.

cuda_memory_info_idx

Query CUDA memory usage for a specific device. Returns (used_bytes, total_bytes) or an error if CUDA is not available.

cuda_peak_active_bytes

Peak bytes allocated to tensors since last cuda_reset_peak_stats() on device 0.

cuda_peak_active_bytes_idx

Peak bytes allocated to tensors since last cuda_reset_peak_stats() on a specific device.

cuda_peak_reserved_bytes

Peak bytes reserved by the CUDA caching allocator since last cuda_reset_peak_stats() on device 0.

cuda_peak_reserved_bytes_idx

Peak bytes reserved by the CUDA caching allocator since last cuda_reset_peak_stats() on a specific device.

cuda_reset_peak_stats

Reset peak memory statistics for device 0.

cuda_reset_peak_stats_idx

Reset peak memory statistics for a specific device. Equivalent to torch.cuda.reset_peak_memory_stats().

cuda_synchronize

Synchronize a CUDA device (wait for all pending work to complete).

cuda_utilization

Query GPU utilization percentage (0-100) via NVML. Returns None if NVML is not available or the query fails.

cuda_utilization_idx

Query GPU utilization percentage for a specific device (0-100) via NVML.

current_cuda_device

Get the current CUDA device index.

hardware_summary

One-line hardware summary for dashboard headers.

live_tensor_count

Number of live C++ Tensor handles (created but not yet dropped). If this grows over time during training, there is a handle leak. If it stays stable but RSS grows, the leak is inside libtorch.

malloc_trim

Ask glibc to return free memory to the OS (Linux only).

manual_seed

Seed all libtorch RNGs (CPU + CUDA) for reproducible tensor ops.

probe_device

Probe whether a CUDA device can execute compute kernels under the current libtorch build. Returns Ok(()) if the device works, or an error describing why it cannot (e.g. missing kernel image for sm_61).

rss_kb

Read current process RSS in kilobytes (Linux only). Returns 0 on non-Linux or if /proc/self/statm is unreadable.

set_cudnn_benchmark

Enable or disable cuDNN benchmark mode.

set_current_cuda_device

Set the current CUDA device.

usable_cuda_devices

Return all CUDA devices that can run compute kernels, with warnings for any excluded devices printed to stderr.

Type Aliases

Result