trnrand

Random number generation for AWS Trainium via NKI.

Seeded pseudo-random distributions, quasi-random sequences for quasi-Monte Carlo, and on-device counter-based RNG (Threefry4x32-20, with a Philox 4x32-10 kernel alongside) on the GpSimd engine.

Part of the trnsci scientific computing suite (github.com/trnsci).

Current phase

trnrand follows the trnsci 5-phase roadmap. Active work is tracked in phase-labeled GitHub issues:

• Phase 1 — correctness (done): Threefry, Box-Muller, Philox, Sobol and Halton NKI kernels are bit-exact against their CPU references on Trn2 (Neuron SDK 2.33, native TorchNeuron).
• Phase 3 — perf (done): wide-tile bulk Threefry kernels (v0.8.0–v0.9.0), opt-in multi-core bulk draws with set_cores (v0.10.0), pinned host output (v0.11.0).
• Phase 4 — multi-chip (done): counter-partitioned generators (partition_rank / partition_size); the counter reuse on P > 1 in 0.7.0–0.11.0 is fixed in v0.12.0.
• Phase 5 — generation (open): trn2 PSUM sizing for larger batched tile generation.

(No Phase 2 for trnrand — the precision story is inherited where relevant.)

Suite-wide tracker: trnsci/trnsci#1.

Known issues in earlier releases

Silent wrong-result bugs fixed in the 2026-10-31 to 2026-11-02 releases. Details are in the CHANGELOG. "NKI" means set_backend("nki"), or "auto" on Neuron hardware.

Install

pip install trnrand

# Trn2/Trn3 with TorchNeuron (native PyTorch, Neuron SDK 2.33+)
pip install "trnrand[torchneuron]" --extra-index-url https://pip.repos.neuron.amazonaws.com

# Trn1 / PyTorch-XLA lineage
pip install "trnrand[neuron]" --extra-index-url https://pip.repos.neuron.amazonaws.com

Kernels dispatch on whichever Neuron PyTorch lineage is installed:

• TorchNeuron (native) — Trn2/Trn3, Neuron SDK 2.33+, Python ≥ 3.12, torch 2.14, torch-neuronx 2.14.*. Each kernel is a PyTorch custom op (torch.ops.trnrand.*) on the neuron device; no torch_xla.
• PyTorch/XLA — Trn1 (and any host on the torch-neuronx 2.9 lineage). Kernels dispatch through torch_xla, unchanged.

trnrand picks the native path when the native torch_neuronx and a neuron device are present, and the XLA path when torch_xla is installed; without nki it uses torch.Generator. Set TRNSCI_NKI_DISPATCH=native or xla (shared by all trnsci libraries) to force a path. With an editable install (pip install -e), set NKI_ENABLE_TRACE_CACHE=0: the NKI kernel cache only versions installed wheels and can serve stale kernels. See installation.

Known platform issue: NKI trace cache and kernel factories

import torch_neuronx turns on the NKI trace cache, and the cache key ignores constants captured in a kernel's closure: two @nki.jit kernels built by one factory, differing only in a captured value, share one trace, and the second silently returns the first's results (aws-neuron/aws-neuron-sdk#1454). trnrand's kernels do not rely on closure constants (or on module globals that can differ between processes): every value that specialises a kernel is a kernel argument, and tests/test_trace_cache_safety.py enforces this. If you write your own NKI kernels with a factory, set NKI_ENABLE_TRACE_CACHE=0 before importing torch_neuronx.

Usage

import trnrand

# Seeded, reproducible generation
g = trnrand.Generator(seed=42)

# Standard distributions
x = trnrand.normal(1000, mean=0.0, std=1.0, generator=g)
u = trnrand.uniform(1000, low=-1.0, high=1.0, generator=g)
e = trnrand.exponential(1000, rate=2.0, generator=g)

# Quasi-random sequences (better convergence for MC integration)
sobol_pts = trnrand.sobol(1024, n_dims=5, seed=42)
halton_pts = trnrand.halton(1024, n_dims=3)
lhs_pts = trnrand.latin_hypercube(100, n_dims=4)

# Module-level seeding
trnrand.manual_seed(42)
x = trnrand.standard_normal(256)

Operations

MC vs QMC Example

python examples/mc_integration.py

Compares pseudo-random vs Sobol quasi-random for estimating the volume of a 5-D hypersphere. QMC converges O(1/N) vs O(1/√N).

Status

• Seeded Generator with state management
• Standard distributions (uniform, normal, exponential, Bernoulli, etc.)
• Sobol, Halton, Latin Hypercube sequences
• MC vs QMC integration example
• NKI Threefry and Philox kernels on GpSimd (bit-exact on Trn2)
• On-device Box-Muller (uniform → normal)
• Counter-partitioned multi-chip generators and multi-core bulk draws (set_cores)
• Benchmarks vs cuRAND

Related Projects

License

Apache 2.0 — Copyright 2026 Scott Friedman

Disclaimer

trnsci is an independent open-source project. It is not sponsored by, endorsed by, or affiliated with Amazon.com, Inc., Amazon Web Services, Inc., or Annapurna Labs Ltd.

"AWS", "Amazon", "Trainium", "Inferentia", "NeuronCore", "Neuron SDK", and related identifiers are trademarks of their respective owners and are used here solely for descriptive and interoperability purposes. Use does not imply endorsement, partnership, or any other relationship.

All work, opinions, analyses, benchmark results, architectural commentary, and editorial judgments in this repository and on trnsci.dev are those of the project's contributors. They do not represent the views, positions, or commitments of Amazon, AWS, or Annapurna Labs.

Feedback directed at the Neuron SDK or Trainium hardware is good-faith ecosystem commentary from independent users. It is not privileged information, is not pre-reviewed by AWS, and should not be read as authoritative about product roadmap, behavior, or quality.

For official AWS guidance, see aws-neuron documentation and the AWS Trainium product page.