feat: add SVHN Quantum Kernel SVM benchmark

[rich7420]

Related Issues

N/A

Changes

• Bug fix
• New feature
• Refactoring
• Documentation
• Test
• CI/CD pipeline
• Other

Why

A Quantum Kernel SVM pipeline eliminates the iterative training loop — encoding becomes ~19% of total pipeline time, making QDP's GPU encoding advantage clearly visible in end-to-end results.

How

Added two new benchmark scripts that implement a Quantum Kernel SVM classification pipeline on SVHN:

SVHN (32×32×3) → StandardScaler → Amplitude Encode (3072→4096, 12 qubits)
  → Quantum Kernel K[i,j] = |⟨ψ(xⱼ)|ψ(xᵢ)⟩|² → SVM (precomputed) → 5-fold CV

New files

• pennylane_baseline/svhn_kernel_amplitude.py — CPU encoding (L2-norm + zero-pad)
• qdp_pipeline/svhn_kernel_amplitude.py — QDP GPU encoding (amplitude)

Benchmark results (RTX 3080)

5000 samples, 5-fold stratified CV, binary classification (digit 1 vs 7), C=100

Step	CPU	QDP GPU	Speedup
Scale + Encode	0.648s	0.342s	1.89×
Kernel (matmul)	1.191s	0.788s	1.51×
CV fit + predict (5 folds)	1.597s	1.170s	1.36×
Total	3.44s	2.30s	1.49×

Accuracy: 0.9104 ± 0.0091 (identical for both pipelines)

How to run

# CPU baseline
python benchmark/encoding_benchmarks/pennylane_baseline/svhn_kernel_amplitude.py

# QDP GPU pipeline
python benchmark/encoding_benchmarks/qdp_pipeline/svhn_kernel_amplitude.py

# Custom parameters
python benchmark/encoding_benchmarks/qdp_pipeline/svhn_kernel_amplitude.py \
  --n-samples 10000 --folds 5 --svm-c 100

SVHN data (~180 MB) is auto-downloaded on first run.

Checklist

• Added or updated unit tests for all changes
• Added or updated documentation for all changes

[ryankert01]

lg, just need to be careful when we want to write an article about it:

• baseline: cpu
• qdp_pipeline: cpu -> gpu(encoding) -> cpu (train)

[ryankert01]

I think in benchmark, we will have several scenarios (eg. 3). We can use different directories to showcase these scenarios, eg.

bench
 -> scenario1
 -> scenario2
 -> scenario3

(can be a follow up to organize these)

current 2 scenario looks like these

1. cpu -> gpu (encode + train)
2. cpu -> gpu (encode) -> cpu (train)

Tho I think there might be better way to split scenarios. These can be further discussed in next community sync Asia.

[guan404ming]

Left some nits, great one. Thank you!

[guan404ming]

Should we remove the pennylane here as well?

[guan404ming]

This is called before time.perf_counter() starts, but there's no prior GPU work to synchronize at that point (the data is still on CPU). This first synchronize is a no-op.