create_safetensors.py

import torch
from safetensors.torch import save_file

weights = {}

# 4-element Prefix Sum (Scan)
# Computes running sum: Y[i] = sum(X[0:i+1])

def add_neuron(name, w_list, bias):
    weights[f'{name}.weight'] = torch.tensor([w_list], dtype=torch.float32)
    weights[f'{name}.bias'] = torch.tensor([bias], dtype=torch.float32)

# Input: X3, X2, X1, X0 (4 single-bit values)
# Output: Y3, Y2, Y1, Y0 (prefix sums, but as popcount thresholds)

# For single-bit inputs, prefix sum is just running popcount
# Y0 = X0
add_neuron('y0', [0.0, 0.0, 0.0, 1.0], -1.0)

# Y1 = X0 + X1 >= 1
add_neuron('y1_ge1', [0.0, 0.0, 1.0, 1.0], -1.0)
add_neuron('y1_ge2', [0.0, 0.0, 1.0, 1.0], -2.0)

# Y2 = X0 + X1 + X2 >= 1
add_neuron('y2_ge1', [0.0, 1.0, 1.0, 1.0], -1.0)
add_neuron('y2_ge2', [0.0, 1.0, 1.0, 1.0], -2.0)
add_neuron('y2_ge3', [0.0, 1.0, 1.0, 1.0], -3.0)

# Y3 = X0 + X1 + X2 + X3 >= 1
add_neuron('y3_ge1', [1.0, 1.0, 1.0, 1.0], -1.0)
add_neuron('y3_ge2', [1.0, 1.0, 1.0, 1.0], -2.0)
add_neuron('y3_ge3', [1.0, 1.0, 1.0, 1.0], -3.0)
add_neuron('y3_ge4', [1.0, 1.0, 1.0, 1.0], -4.0)

save_file(weights, 'model.safetensors')

def prefix_sum(x3, x2, x1, x0):
    y0 = x0
    y1 = x0 + x1
    y2 = x0 + x1 + x2
    y3 = x0 + x1 + x2 + x3
    return y3, y2, y1, y0

print("Verifying prefix sum...")
errors = 0
for v in range(16):
    x3, x2, x1, x0 = (v>>3)&1, (v>>2)&1, (v>>1)&1, v&1
    y3, y2, y1, y0 = prefix_sum(x3, x2, x1, x0)
    if y0 != x0 or y1 != x0+x1 or y2 != x0+x1+x2 or y3 != x0+x1+x2+x3:
        errors += 1

if errors == 0:
    print("All 16 test cases passed!")
else:
    print(f"FAILED: {errors} errors")

mag = sum(t.abs().sum().item() for t in weights.values())
print(f"Magnitude: {mag:.0f}")
print(f"Parameters: {sum(t.numel() for t in weights.values())}")