chore: expose p-error to cifar-16b benchmark

.github/workflows/cifar_benchmark.yaml

@@ -26,6 +26,11 @@ on:
         default: "3"
         type: string
         required: true
+      p_error:
+        description: P-error to use
+        default: "0.01"
+        type: string
+        required: true
 
   # FIXME: Add recurrent launching
   # https://github.com/zama-ai/concrete-ml-internal/issues/1851
@@ -128,7 +133,7 @@ jobs:
         if: github.event.inputs.benchmark == 'cifar-10-16b'
         run: |
           source .venv/bin/activate
-          NUM_SAMPLES=${{ github.event.inputs.num_samples }} python3 ./use_case_examples/cifar/cifar_brevitas_training/evaluate_one_example_fhe.py
+          NUM_SAMPLES=${{ github.event.inputs.num_samples }} P_ERROR=${{ github.event.inputs.p_error }} python3 ./use_case_examples/cifar/cifar_brevitas_training/evaluate_one_example_fhe.py
           python3 ./benchmarks/convert_cifar.py --model-name "16-bits-trained-v0"
 
       - name: Archive raw predictions

use_case_examples/cifar/cifar_brevitas_training/evaluate_one_example_fhe.py

@@ -20,6 +20,7 @@
 # FIXME: https://github.com/zama-ai/concrete-ml-internal/issues/3953
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 NUM_SAMPLES = int(os.environ.get("NUM_SAMPLES", 1))
+P_ERROR = float(os.environ.get("P_ERROR", 0.01))
 
 
 def measure_execution_time(func):
@@ -83,7 +84,7 @@ def wrapper(*args, **kwargs):
 print("Compiling the model.")
 quantized_numpy_module, compilation_execution_time = measure_execution_time(
     compile_brevitas_qat_model
-)(torch_model, x, configuration=configuration, rounding_threshold_bits=6, p_error=0.01)
+)(torch_model, x, configuration=configuration, rounding_threshold_bits=6, p_error=P_ERROR)
 assert isinstance(quantized_numpy_module, QuantizedModule)
 
 print(f"Compilation time took {compilation_execution_time} seconds")