fix: update QDP encoding method and synchronize CUDA device in SVHN IQP scripts

Author: ryankert01

qdp/qdp-python/benchmark/encoding_benchmarks/README.md

@@ -148,14 +148,16 @@ uv run python benchmark/encoding_benchmarks/pennylane_baseline/svhn_iqp.py \
 ## SVHN IQP (QDP pipeline)
 
 Pipeline is identical to the baseline except for encoding:
-PCA-reduced vectors → QDP `QdpEngine.encode(method="iqp")` (one-shot, GPU) → `StatePrep(state_vector)` → same variational classifier.
+PCA-reduced vectors → QDP `QdpEngine.encode(encoding_method="iqp")` (one-shot, GPU) → `StatePrep(state_vector)` → same variational classifier.
 
 ```bash
 uv run python benchmark/encoding_benchmarks/qdp_pipeline/svhn_iqp.py
 ```
 
 The CLI mirrors the baseline, plus:
 
+> **Note:** The QDP pipeline always performs the encoding step on a CUDA GPU via QDP. A CUDA-capable device is required even when you select `--backend cpu` for the training backend.
+
 - **QDP-specific flags**
   - `--device-id`: CUDA device id (default: 0)
 
@@ -166,15 +168,9 @@ uv run python benchmark/encoding_benchmarks/qdp_pipeline/svhn_iqp.py \
   --n-qubits 6 --n-samples 200 --iters 200 --trials 1 --early-stop 0 --backend cpu
 ```
 
-## SVHN IQP experiment runner
-
-Run all qubit-scaling, accuracy-parity, and sample-scaling experiments:
-
-```bash
-bash benchmark/encoding_benchmarks/run_svhn_iqp_experiments.sh
-```
+## SVHN IQP experiments
 
-Logs are saved to `benchmark/encoding_benchmarks/logs/`. Results and analysis are in `benchmark/encoding_benchmarks/report.md`.
+Experiment logs are saved to `benchmark/encoding_benchmarks/logs/`.
 
 ## Full help
 

qdp/qdp-python/benchmark/encoding_benchmarks/qdp_pipeline/svhn_iqp.py

@@ -606,11 +606,11 @@ def main() -> None:
     X_train_params = features_to_iqp_params(X_train_pca, n_qubits)
     X_test_params = features_to_iqp_params(X_test_pca, n_qubits)
 
-    torch.cuda.synchronize()
+    torch.cuda.synchronize(args.device_id)
     t0 = time.perf_counter()
     encoded_train = encode_qdp(X_train_params, n_qubits, args.device_id)
     encoded_test = encode_qdp(X_test_params, n_qubits, args.device_id)
-    torch.cuda.synchronize()
+    torch.cuda.synchronize(args.device_id)
 
     if args.backend == "gpu":
         # Zero copy: QDP CUDA tensor -> lightning.gpu StatePrep (stays on GPU)