benchmark: make runs deterministic and reproducible

benchmarks/README.md

@@ -45,6 +45,20 @@ uv run --extra benchmarks python benchmarks/regression_benchmark.py
    python benchmarks/regression_benchmark.py
    ```
 
+## Reproducibility
+
+Both scripts fix every random seed and restrict PyTorch to deterministic kernels, so repeated runs on the same machine and software stack produce identical results.
+
+On CUDA, one more setting is needed and has to come from the environment:
+
+```bash
+export CUBLAS_WORKSPACE_CONFIG=:4096:8
+```
+
+Without it, `torch.use_deterministic_algorithms(True)` raises as soon as a cuBLAS reduction runs.
+
+Results still differ across GPU architectures, CUDA/cuDNN versions and PyTorch releases: floating-point addition is not associative, so a different kernel sums in a different order. The numbers reported here were produced on the hardware listed under [Runtime](#runtime); expect small deviations elsewhere.
+
 ## Requirements
 
 - **Python:** 3.10–3.13

benchmarks/classification_benchmark.py

@@ -48,6 +48,13 @@ def set_seed(seed=42):
         torch.cuda.manual_seed_all(seed)
 
 
+def enable_determinism():
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+    torch.use_deterministic_algorithms(True)
+
+
+enable_determinism()
 set_seed(42)
 print(f"Running Bensemble Benchmark on {DEVICE}")
 

benchmarks/regression_benchmark.py

@@ -2,6 +2,7 @@
 Regression benchmark: PBP vs VI (ELBO) vs Laplace (K-FAC) vs MAP baseline on UCI datasets.
 """
 
+import random
 import time
 from pathlib import Path
 
@@ -25,6 +26,7 @@
     f"Running Bensemble Regression Benchmark on {DEVICE} (PBP itself always runs on CPU — see run_pbp)"
 )
 
+SEED = 42
 HIDDEN = 50
 N_SPLITS = 5
 TEST_FRACTION = 0.1
@@ -36,6 +38,23 @@
 RESULTS_DIR = Path("results_regression")
 RESULTS_DIR.mkdir(exist_ok=True)
 
+
+def set_seed(seed=SEED):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+def enable_determinism():
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+    torch.use_deterministic_algorithms(True)
+
+
+enable_determinism()
+
 DATASETS = {
     "yacht": {"uci_id": None, "target_col": None, "direct": True},
     "energy": {"uci_id": 242, "target_col": "Y1"},
@@ -314,6 +333,7 @@ def main():
                     Xtr_raw, ytr_raw, Xte_raw, yte_raw
                 )
 
+                set_seed(SEED + split_idx)
                 mean, var, elapsed = runner(Xtr, ytr, Xte, yte, epochs=NUM_EPOCHS)
                 rmse, nlpd = regression_metrics(mean, var, yte, y_mean, y_std)