[QCS] 🌠 Wander/漫步者: Issue #33 Compile VQE Forward, Reverse, and Ansatz with VQETape

[JunkaiWang-TheoPhy]

The efficient move is not the one that avoids the board, but the one that finally understands it.

高效的一步并非绕开棋盘，而是终于看懂棋盘。

第二幕 · 让计算穿过物质的风暴

第十章 · 看懂棋盘，而不是绕开它

VQE 的代价从不只藏在一枚棋子里。前向收缩、反向传播、内存、路径与 ansatz 彼此改变，

高效的一步来自对整张棋盘的理解。

← 上一章：同一阵风吹向两种手性     ·     下一幕：两道阴影之间的窄光 →

Team

Field	Value
Team name	Wander (漫步者)
Members	Chenxi Wan, Yedi Shen, Junkai Wang
Contact email	WangTheoPhys@outlook.com

VQETape: differentiated co-design for exact VQE

Compile the forward contraction, reverse program, and variational ansatz as one optimization problem.

Team	Wander (漫步者)
Members	Chenxi Wan, Yedi Shen, Junkai Wang
Contact email	WangTheoPhys@outlook.com
Track	Quantum circuit simulation (qcs)
Challenge	Addresses #33, “Extreme-efficiency VQE numerical simulation,” released by Shi-Xin Zhang
Public showcase	JunkaiWang-TheoPhy/issue-33-extreme-efficiency-vqe

Reviewer package

• Seven-page technical report (PDF)
• Technical report (Markdown)
• Standalone HTML report
• Matched RTX 3090 benchmark TSV
• Result map
• Structured report JSON
• SHA256 artifact manifest

The package is generated from six committed canonical JSON records:

cd tracks/qcs/solutions/WangTheoPhys/issue33
python scripts/build_submission_report.py

Demonstrated same-machine threshold

All rows use the same open-boundary TFIM workload, RTX 3090 node, seed, precision policy, energy, complete gradient, and synchronization boundaries. The declared objective is compile + first + 100 * warm median.

Implementation	Compile (s)	First (s)	Warm (ms)	Objective (s)	Host RSS (MiB)	NVML (MiB)
TensorCircuit-NG / OMECo	17.7799	0.2263	2.6577	18.2720	661.3	272
VQETape / statevector	28.7213	0.4025	3.3785	29.4617	462.4	272
VQETape / direct TN	45.6544	0.1376	5.0498	46.2970	575.0	274
VQETape / spatial block-2	15.5816	0.3636	8.3281	16.7781	473.9	274

The selected VQETape spatial program improves the controlled 100-step objective by 8.2% and reduces host peak RSS by 28.3%. TensorCircuit-NG's 2.6577 ms steady-state point anchors the next warm-kernel fusion objective; VQETape statevector currently records 3.3785 ms. Job-level device samples form a measured 272–274 MiB range.

Four innovations

1. Differentiated contraction programming

VQETape serializes the contraction tree, constructs algebraic transpose einsums, and measures backward FLOPs, traffic, live residuals, largest intermediates, and checkpoint choices. The reverse program becomes a selectable compiler object alongside the forward path.

2. Exact spatial-transfer lowering

The bra–MPO–ket network is lowered into first, repeated bulk, optional tail, and last programs. Only the exact boundary is carried; block width, scan, rematerialization, segmented adjoints, explicit VJPs, and guarded Z2 compression enter one search.

3. Commutator-complete adaptive ansatz

Exact insertion gradients and Fubini–Study normalization rank YZ/ZY candidates together with contraction-cost deltas. The adaptive 10-parameter circuit reaches 5.05e-11 energy error; the fixed 14-parameter control records 1.70e-7 under the same audited budget.

4. Correctness-gated auto-evaluation

Fresh processes capture compile, first, warm, host RSS, compiler memory, logical residuals, device samples, energy, and the complete gradient. A precision bridge maps the declared JAX policy into TensorNetwork's cached backend, making numerical precision an explicit reproducibility contract.

Relation to TensorCircuit-NG

TensorCircuit-NG provides a unified tensor-native computational graph, AD/JIT, slicing, and distributed execution. Its contractor stack optimizes forward time, space, and read/write objectives with OMECo and related search tools.

VQETape contributes the differentiated-program layer above that foundation: reverse contraction programs, saved-residual schedules, exact physics reductions, optimizer state, and ansatz growth share one correctness gate and measured cost vector. Joint co-design makes cross-layer exchanges directly searchable—for example, spending forward work to shrink a live tape or selecting a gate for both energy descent and boundary cost.

Validated Fig. 2 construction

The separate SU(4) runner preserves 15 * L * (N - 1) parameters, TensorNetwork FiniteTFI MPO construction, find/execute separation, slicing controls, and checksum-bound JSON path artifacts. An RTX 3080 execution at N=6,L=3 records 2.38e-7 energy error and 3.29e-7 gradient relative L2 error. The paper-comparable N=32,L=16 execution is the declared scale-up target.

Verification

• fresh full regression: 406 tests passed, with six declared structural cases;
• targeted TensorCircuit-NG baseline and Fig. 2 suite: 17 tests passed;
• six canonical report inputs and 27 total evidence JSON files parse;
• all src/vqetape modules compile;
• 7-page A4 PDF passed text extraction and page-by-page visual QA;
• TSV, SHA256 manifest, git diff --check, and secret-pattern scans completed.

Reproduce

cd tracks/qcs/solutions/WangTheoPhys/issue33
python3.12 -m venv .venv
.venv/bin/python -m pip install -e '.[test,baseline]'
.venv/bin/python -m pytest -q
python scripts/build_submission_report.py

The result is a reproducible exact-VQE compiler platform with two instrumented next targets: warm-kernel fusion and the validated N=32,L=16 deployment.

[JunkaiWang-TheoPhy]

RTX 3090 precision-aware co-design milestone

The same-machine GPU study now establishes three exact VQETape execution representations under one value-gradient contract:

Representation	Compile	First	Warm median	Compile + first + 100 warm	Host RSS
Statevector	28.7213 s	0.4025 s	3.3785 ms	29.4617 s	462.4 MiB
Direct TN / dense MPO	45.6544 s	0.1376 s	5.0498 ms	46.2970 s	575.0 MiB
Spatial transfer / block 2	15.5816 s	0.3636 s	8.3281 ms	16.7781 s	473.9 MiB

The controlled TensorCircuit-NG / OMECo reference records 18.2720 s on the declared 100-step objective, 2.6577 ms warm execution, and 661.3 MiB host RSS. VQETape spatial therefore demonstrates an 8.2% objective improvement and 28.3% host RSS reduction; the TensorCircuit-NG warm point defines the next fusion target.

The precision bridge is part of the compiler contract: fresh workers declare JAX_DEFAULT_MATMUL_PRECISION=highest, and the TensorNetwork cached backend receives the matching precision policy. The validated spatial execution records 3.815e-6 energy error and 9.145e-7 gradient relative L2 error.

This milestone turns representation, differentiated program, precision policy, and horizon-aware selection into one auditable search surface. Canonical JSON, SHA256 provenance, and the regenerated report package are linked from the PR body.

[JunkaiWang-TheoPhy]

@fliingelephant VQETape’s complete Issue #33 delivery is ready for review.

The central contribution is a differentiated co-design compiler for exact VQE: it compiles the forward contraction, reverse program, saved-residual/checkpoint schedule, optimizer, and variational ansatz as one measured optimization problem.

Reviewer entry points

• Public self-contained repository
• Seven-page technical report (PDF)
• Searchable technical report
• Matched benchmark data
• Delivery result map
• Implementation and tests

Demonstrated anchors

• 8.2% improvement on compile + first + 100 warm versus the controlled TensorCircuit-NG / OMECo reference;
• 28.3% reduction in host peak RSS on the same RTX 3090 workload;
• 5.05e-11 adaptive-ansatz energy error with 10 parameters;
• validated TensorCircuit-NG Fig. 2 construction with direct energy and complete-gradient comparison;
• 406-test fresh full regression plus 17 targeted baseline/Fig. 2 tests.

The measured 2.6577 ms TensorCircuit-NG warm reference and validated N=32,L=16 protocol now define two concrete next objectives. All reviewer artifacts are generated from committed canonical JSON and bound by SHA256.