Skip to content

Backends

nexuml backend list is the runtime catalog for several independent extension points. "Backend" is not one universal interface; each category belongs to a different concern.

nexuml backend list
nexuml backend list data-loader
nexuml backend list data-export

Catalog vs dependency health

The command lists registered backend definitions/implementations. A listed optional backend can still require a third-party package at materialization time. For example, DaliLoader is registered even when nvidia.dali is not importable.

Data loaders

Name Role
torch standard PyTorch loading
dali NVIDIA DALI loader; optional platform-specific runtime
tensor_shards windowed loading from materialized tensor shards

See Data loading.

Dataset export

Built-in data-export names currently include:

  • numpy
  • numpy_mmap
  • torch
  • tensordict_memmap
  • webdataset
  • tensor_shards

See Dataset export.

Training / execution

The catalog exposes lightning and ray in the training category. Conceptually, Lightning remains the canonical session/training lifecycle and Ray changes distributed placement/setup around that lifecycle.

See Execution modes.

Tracking

The catalog includes TensorBoard, DVCLive, and MLflow integrations. Their dependencies/configuration are documented in Tracking and logging.

Pipeline export

The catalog includes:

  • package — primary NexuML train-package artifact;
  • safetensors — weight export;
  • onnx — inference graph export for supported pipelines.

See Export and reload.

Evaluation storage

Distance-estimator feature accumulation uses ram or memmap. DistanceEstimatorSpec.create_feature_store() constructs that configured store and applies its path, sample limit, and retention settings.

TensorDict temporary storage is a separate contract that uses memory or memmap. ram and memory are intentionally not aliases: each belongs to a different storage family.

The generated Python API is the exact reference for the currently installed version.