trntensor

Tensor contractions for AWS Trainium via NKI.

Einstein summation with contraction planning, CP and Tucker decompositions. Expresses scientific tensor workloads naturally instead of decomposing to GEMM. Part of the trnsci scientific computing suite (github.com/trnsci).

Current phase

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

• Phase 1 — correctness (closed): NKI kernels validated on hardware. On the native TorchNeuron path (Trn2/Trn3) the fused mp2_energy, ao_to_mo_transform, df_exchange and df_coulomb kernels beat the aten routes they replace; the NKI matmul/bmm kernels never did and are deprecated there in favour of aten (0.19.0).
• Phase 2 — precision (closed): precision-aware mode selection (precision=, target_forward_error=). precision="dd" on Trainium still waits on trnblas#22 double-double GEMM.
• Phase 3 — perf (closed): greedy multi-operand path planner, plan cache.
• Phase 4 — multi-chip (active): sharded tensor contractions (ShardedTensor, reduce-parallel all-reduce).
• Phase 5 — generation: trn2 fused multi-contraction paths.

Suite-wide tracker: trnsci/trnsci#1.

Install

pip install trntensor

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

# Trn1 / PyTorch-XLA lineage
pip install "trntensor[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.trntensor.*) 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.

trntensor picks the native path when the native torch_neuronx and a neuron device are present, and the XLA path when torch_xla is installed; otherwise it runs on PyTorch. Set TRNSCI_NKI_DISPATCH=native or xla (shared by all trnsci libraries) to force a path. Pin operands on the device with to_device / from_device; the older to_xla / from_xla still work on both lineages (see the migration note). 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.

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). trntensor'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 torch
import trntensor

# Einsum — drop-in for torch.einsum with contraction planning
C = trntensor.einsum("ij,jk->ik", A, B)  # matmul
T = trntensor.einsum("ap,bp->ab", B_i, B_j)  # DF-MP2 contraction
X = trntensor.einsum("mi,mnP->inP", C_occ, eri)  # AO→MO transform

# Contraction planning
plan = trntensor.plan_contraction("ij,jk->ik", A, B)
flops = trntensor.estimate_flops("ij,jk->ik", A, B)

# CP decomposition (tensor hypercontraction)
factors, weights = trntensor.cp_decompose(tensor, rank=10)
reconstructed = trntensor.cp_reconstruct(factors, weights)

# Tucker decomposition (HOSVD)
core, factors = trntensor.tucker_decompose(tensor, ranks=(5, 5, 5))

Operations

Category	Operation	Description
Contraction	einsum	General tensor contraction
Contraction	multi_einsum	Multiple contractions (fusion-ready)
Planning	plan_contraction	Analyze and select strategy
Planning	estimate_flops	FLOPs for a contraction
Decomposition	cp_decompose	CP/PARAFAC via ALS
Decomposition	tucker_decompose	Tucker via HOSVD
Decomposition	tt_decompose	Tensor Train via TT-SVD
Chemistry	mp2_energy	Fused DF-MP2 correlation energy
Chemistry	ao_to_mo_transform	Fused AO→MO transform (full or packed integrals)
Chemistry	df_exchange / df_coulomb	Fused DF exchange (K) / Coulomb (J) from packed integrals
Chemistry	pack_tril / unpack_tril	Packed (μ≥ν) three-centre integral layout
Sharding	ShardedTensor, scatter, gather	Output- and reduce-parallel sharded contractions

Status

• Einsum with matmul/bmm/torch dispatch (output index order honoured since 0.22.0; earlier releases returned the wrong layout for transposed/permuted outputs, #64)
• Contraction planner
• CP decomposition (ALS)
• Tucker decomposition (HOSVD)
• DF-MP2 einsum example
• NKI fused contraction kernels (mp2_energy, ao_to_mo_transform, df_exchange, df_coulomb)
• Device operand residency (to_device / from_device; to_xla / from_xla kept)
• Native TorchNeuron dispatch (Trn2/Trn3, torch.ops.trntensor.*) alongside PyTorch/XLA (Trn1)
• NKI CPU simulator + nki-simulator CI gate
• Optimal contraction ordering — greedy path search for 3+ operands
• Multi-contraction shared-operand XLA pre-pinning
• Contraction plan cache (clear_plan_cache / plan_cache_info)
• Alpha/beta scaling for einsum (cuTENSOR GEMM-style)
• Input validation with descriptive errors
• Tensor Train (TT) decomposition (TT-SVD)
• Non-negative CP + warm-start CP
• PEP 561 py.typed marker
• Mixed precision / dtype override for bf16/fp16
• precision= kwarg ("fast" / "kahan" fp64 promotion / "sr" / "dd": CPU emulation only, raises on Trainium until trnblas#22)

Known issues in earlier releases

Silent wrong-result bugs fixed in the 2026-10-31 to 2026-11-04 releases. Details are in the CHANGELOG.

Bug	Affected	Fixed in
einsum returned the wrong layout for transposed or permuted outputs (e.g. ij,jk->ki gave A@B), every backend (#64)	through 0.21.0	0.22.0
einsum(..., out=C) with beta=0 read C, so NaN/inf in it made the result NaN (#71)	0.3.0–0.22.0	0.23.0
tucker_decompose used Uᵀ instead of Uᴴ for complex tensors, giving a wrong core (#73)	through 0.22.0	0.23.0
precision="kahan" / "dd" / target_forward_error dropped the imaginary part of complex operands and rounded large integers (#69)	"kahan" 0.4.0–0.22.0; "dd", target_forward_error 0.17.0–0.22.0	0.23.0
einsum(precision="dd") on float32/float64 was less accurate than "fast" (CPU emulation, #84)	0.17.0–0.22.0	0.23.0
multi_einsum returned float32 results as float64 in mixed-dtype batches (#75)	0.17.0–0.22.0	0.23.0
NKI chemistry kernels returned inf for float64 operands, and some kernel routes changed the result dtype (#87)	df_* 0.21.0–0.22.0; ao_to_mo_transform every release with its kernels	0.23.0
ao_to_mo_transform aten route on the TorchNeuron device was wrong once eri exceeded 4 GiB (#97; the LNC=1 gather case is the 0.24.0 row below)	through 0.22.0	0.23.0
unpack_tril / pack_tril (and the packed-eri aten / legacy routes of ao_to_mo_transform) on the neuron device at LNC=1 returned wrong data, at times NaN, once one gathered output reached 4 GiB (154 M mismatched elements at 4.99e9 bytes) (#97)	0.20.0–0.23.0	0.24.0
Complex tucker_decompose on the neuron device used Uᵀ, not Uᴴ: device matmul ignores the lazy conj bit of .mH (#89)	0.23.0	0.24.0
float64 on the neuron device rounded to fp32 in gather (device torch.cat) and in pack_tril / unpack_tril / pack_density / unpack_packed (device index_select / fancy indexing), rel. err ~4e-8 (#92)	through 0.23.0	0.24.0
multi_einsum under "auto" pinned shared host float64 / complex128 operands to the device, where device aten computed them at fp32 precision (#92)	through 0.23.0	0.24.0
Integer, bool, complex and mixed-dtype operands of einsum / multi_einsum reached the fp32/bf16/fp16 matmul kernels under set_backend("nki"), or under "auto" above the size thresholds (trnsci/trnsci#35)	through 0.23.0	0.24.0
einsum(..., out=) never wrote out, so the documented einsum(..., beta=1.0, out=acc) accumulation left acc unchanged (#77)	0.3.0–0.23.0	0.24.0
cp_reconstruct returned float32 for bf16 / fp16 factors (#78)	0.1.1–0.23.0	0.24.0

Related Projects

Project	What
trnfft	FFT + complex ops
trnblas	BLAS operations
trnsolver	Linear solvers
trnrand	Random number generation
trnsparse	Sparse operations

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.