[lkdvos]

ncon spent most of its time on bookkeeping rather than on contractions: runtime type inference (promote_op) at every node of the contraction tree, abstractly typed containers, repeated keyword argument handling, and unspecialized network analysis. This PR makes the label handling type stable, so that every pairwise contraction goes through a single dynamic dispatch. It also includes the fix for #273, previously #315.

• Type-stable BLAS contraction. makeblascontractable received the scalar type as an unconstrained argument, which Julia does not specialize on, so the returned tensor was inferred as Any and the rest of the BLAS contraction path was dynamically dispatched. It also returned either its input or a permuted copy, whose types differ when the copy comes from an allocator such as Bumper's, and that union was heap allocated on every contraction. makeblascontractable is removed: _blas_contract! now makes the inputs BLAS-contractable one at a time and passes them on within each branch, so the union is never materialized. Each input is reshaped into a matrix before it is passed on, so the later steps do not depend on its number of indices, which keeps the number of compiled methods down. This also speeds up @tensor (see the reference row below).
• Specialize the network analysis. ncontree.jl and indexordertree.jl were included in the module-level @nospecialize block meant for the macros, but are also used at runtime by ncon.
• Faster validation and tree construction. Labels are counted rather than repeatedly searched and deleted, and the contraction tree is built in a type-stable manner as a list of pairwise contractions. The nested trees returned by ncontree and indexordertree are unchanged, except that disconnected parts of the network are now combined from left to right.
• Type-stable ncon. Tensors, labels, output and conjugation flags are canonicalized once, all label manipulations happen on Vector{Int}, and keyword arguments are processed once.
• contract_indices for labels stored in vectors. This avoids the type-unstable tuple implementation, which is compiled for every combination of numbers of indices. It also applies to tensorcontract(IC, A, IA, B, IB) with vector labels.

As a side effect, traces can now be specified with row matrices as index lists.

Fixes #273, together with QuantumKitHub/Strided.jl#94. Allocations in its reproducer (StridedBLAS, Bumper allocator, latest Strided and StridedViews main):

	main	this PR	this PR + Strided.jl#94
Float64	112 B	48 B	0 B
ComplexF64	112 B	48 B	0 B
ComplexF64, conj	48 B	48 B	0 B

Benchmarks

Exclusive node, Julia 1.13.1, a single thread, every measurement in a fresh process, registered Strided. Tensors have dimension 2 per index unless noted, so the overhead dominates. Minimum times; three interleaved rounds agree to within 2%.

	main	this PR
5 tensors, mixed arity	58.8 μs / 386 allocs	6.3 μs / 149 allocs
4 tensors, mixed arity, Tuple input	44.2 μs / 290 allocs	4.8 μs / 126 allocs
matrix chain (4)	28.5 μs / 228 allocs	3.0 μs / 115 allocs
matrix chain, order keyword	33.5 μs / 361 allocs	2.9 μs / 109 allocs
with trace	9.0 μs / 85 allocs	3.0 μs / 68 allocs
10-tensor MPS ladder, ComplexF64	137.9 μs / 842 allocs	19.8 μs / 321 allocs
MPS transfer matrix, D = 32	108.0 μs / 311 allocs	58.4 μs / 126 allocs
@tensor reference (first row)	3.47 μs / 46 allocs	2.52 μs / 22 allocs
#273 reproducer, Bumper allocator	0.57 μs / 112 B	0.57 μs / 48 B
same, default allocator	2.35 μs / 16 allocs	0.61 μs / 4 allocs
8⁴ tensors, permuted copies needed	18.2 μs / 25 allocs	14.6 μs / 7 allocs
first ncon call / with new types	2.46 s / 1.18 s	1.74 s / 0.69 s
first calls of 12 networks with different numbers of indices	7.1 s	4.4 s

The last row is from an earlier run with the same setup (the previous version of this PR, without the #273 fix); the final version matches it within noise in local comparisons (median 4.32 s without and 4.39 s with the fix).

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 96.13260% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/indexnotation/ncontree.jl	90.47%	6 Missing ⚠️
src/implementation/indices.jl	96.00%	1 Missing ⚠️

Files with missing lines	Coverage Δ
ext/TensorOperationsChainRulesCoreExt.jl	83.52% <ø> (ø)
src/TensorOperations.jl	100.00% <ø> (ø)
src/implementation/blascontract.jl	100.00% <100.00%> (ø)
src/implementation/ncon.jl	100.00% <100.00%> (ø)
src/indexnotation/indexordertree.jl	100.00% <100.00%> (+14.00%)	⬆️
src/implementation/indices.jl	98.09% <96.00%> (-0.72%)	⬇️
src/indexnotation/ncontree.jl	90.21% <90.47%> (-2.21%)	⬇️

... and 5 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Jutho]

Not sure if we still need this. I also don't understand the Zygote comment.

[lkdvos]

The main point is that various constructions are not compatible with Zygote because internally they do some mutation and have no rules. The comment wasn't very clear and the functions weren't either, so I think I now have a new implementation that is better and bypasses most of this?

[Jutho]

Thanks for this PR. Definitely contains several good improvements, but quite a big PR with also some changes that I don't (yet?) understand the benefit of and seem to make the code more complicated to understand.

[lkdvos]

Alright, I think the comments have been addressed here, thanks for the careful review, I agree I should have gone through the tree manipulation implementations a bit more carefully but I am not too familiar with that piece of the code so I definitely missed some obvious simplifications. Let me know what you think about this version, happy to still adapt some things!