trnsci 0.1.0

Scientific computing suite for AWS Trainium via NKI (FFT, BLAS, RAND, SOLVER, SPARSE, TENSOR)

trnsci

Scientific computing suite for AWS Trainium via NKI. Six focused libraries providing the cu* equivalents the Neuron SDK doesn't ship.

Suite

Project	NVIDIA analog	Scope
trnfft	cuFFT	FFT, complex tensors, complex NN layers
trnblas	cuBLAS	Level 1–3 BLAS, batched GEMM
trnrand	cuRAND	Philox PRNG, Sobol/Halton QMC, LHS
trnsolver	cuSOLVER	Cholesky/LU/QR, Jacobi eigh, CG/GMRES
trnsparse	cuSPARSE	CSR/COO, SpMV/SpMM, Schwarz screening
trntensor	cuTENSOR	einsum with contraction planning, CP/Tucker

Install

# Meta-package with all six sub-projects
pip install trnsci[all]

# Individual components
pip install trnsci[fft]     # just trnfft
pip install trnsci[blas]    # just trnblas
# ... etc

# On Neuron hardware
pip install trnsci[all,neuron]

Development install

git clone [e-mail]:trnsci/trnsci.git
cd trnsci
make install-dev   # pip install -e on each sub-project + umbrella
make test-all      # run pytest across all sub-projects

Cross-project example

python examples/quantum_chemistry/df_mp2_synthetic.py --demo

DF-MP2 energy evaluation composing trnblas (half-transforms), trnsolver (Cholesky of DF metric), and trntensor (einsum contraction). See also examples/nvidia_samples/ for direct ports of canonical NVIDIA CUDA samples.

Status

All sub-projects are Alpha. CPU/PyTorch fallback is functional end-to-end. NKI kernels are scaffolded across the suite; on-hardware validation is the next milestone.

Community

• Site: trnsci.dev
• Contributing guidelines
• Code of Conduct

License

Apache 2.0 — Copyright 2026 Scott Friedman