create_safetensors.py

import torch
from safetensors.torch import save_file

# Min of two 2-bit unsigned numbers
# Inputs: a1, a0, b1, b0
# Outputs: m1, m0 = min(a, b)
#
# Logic: if a <= b then output a, else output b
# Same comparison as max, but swap output selection

weights = {}

# Layer 1: Basic comparisons (same as max2)
weights['l1.a1_gt_b1.weight'] = torch.tensor([[1.0, 0.0, -1.0, 0.0]], dtype=torch.float32)
weights['l1.a1_gt_b1.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l1.b1_gt_a1.weight'] = torch.tensor([[-1.0, 0.0, 1.0, 0.0]], dtype=torch.float32)
weights['l1.b1_gt_a1.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l1.a0_gt_b0.weight'] = torch.tensor([[0.0, 1.0, 0.0, -1.0]], dtype=torch.float32)
weights['l1.a0_gt_b0.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l1.b0_gt_a0.weight'] = torch.tensor([[0.0, -1.0, 0.0, 1.0]], dtype=torch.float32)
weights['l1.b0_gt_a0.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l1.both1_high.weight'] = torch.tensor([[1.0, 0.0, 1.0, 0.0]], dtype=torch.float32)
weights['l1.both1_high.bias'] = torch.tensor([-2.0], dtype=torch.float32)

weights['l1.both1_low.weight'] = torch.tensor([[-1.0, 0.0, -1.0, 0.0]], dtype=torch.float32)
weights['l1.both1_low.bias'] = torch.tensor([0.0], dtype=torch.float32)

weights['l1.a1.weight'] = torch.tensor([[1.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l1.a1.bias'] = torch.tensor([-0.5], dtype=torch.float32)
weights['l1.a0.weight'] = torch.tensor([[0.0, 1.0, 0.0, 0.0]], dtype=torch.float32)
weights['l1.a0.bias'] = torch.tensor([-0.5], dtype=torch.float32)
weights['l1.b1.weight'] = torch.tensor([[0.0, 0.0, 1.0, 0.0]], dtype=torch.float32)
weights['l1.b1.bias'] = torch.tensor([-0.5], dtype=torch.float32)
weights['l1.b0.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0]], dtype=torch.float32)
weights['l1.b0.bias'] = torch.tensor([-0.5], dtype=torch.float32)

# Layer 2
weights['l2.a1_eq_b1.weight'] = torch.tensor([[0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l2.a1_eq_b1.bias'] = torch.tensor([-1.0], dtype=torch.float32)

for v in ['a1_gt_b1', 'b1_gt_a1', 'a0_gt_b0', 'b0_gt_a0', 'a1', 'a0', 'b1', 'b0']:
    idx = ['a1_gt_b1', 'b1_gt_a1', 'a0_gt_b0', 'b0_gt_a0', 'both1_high', 'both1_low', 'a1', 'a0', 'b1', 'b0'].index(v)
    w = [0.0] * 10
    w[idx] = 1.0
    weights[f'l2.{v}.weight'] = torch.tensor([w], dtype=torch.float32)
    weights[f'l2.{v}.bias'] = torch.tensor([-0.5], dtype=torch.float32)

# Layer 3
weights['l3.a_gt_b_part2.weight'] = torch.tensor([[1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l3.a_gt_b_part2.bias'] = torch.tensor([-2.0], dtype=torch.float32)

weights['l3.a0_neq_b0.weight'] = torch.tensor([[0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l3.a0_neq_b0.bias'] = torch.tensor([-1.0], dtype=torch.float32)

for v in ['a1_gt_b1', 'a1', 'a0', 'b1', 'b0', 'a1_eq_b1']:
    if v == 'a1_eq_b1':
        idx = 0
    else:
        idx = ['a1_eq_b1', 'a1_gt_b1', 'b1_gt_a1', 'a0_gt_b0', 'b0_gt_a0', 'a1', 'a0', 'b1', 'b0'].index(v)
    w = [0.0] * 9
    w[idx] = 1.0
    weights[f'l3.{v}.weight'] = torch.tensor([w], dtype=torch.float32)
    weights[f'l3.{v}.bias'] = torch.tensor([-0.5], dtype=torch.float32)

# Layer 4
weights['l4.a_gt_b.weight'] = torch.tensor([[1.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l4.a_gt_b.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l4.a_eq_b.weight'] = torch.tensor([[0.0, -1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0]], dtype=torch.float32)
weights['l4.a_eq_b.bias'] = torch.tensor([-1.0], dtype=torch.float32)

for v in ['a1', 'a0', 'b1', 'b0']:
    idx = ['a_gt_b_part2', 'a0_neq_b0', 'a1_gt_b1', 'a1', 'a0', 'b1', 'b0', 'a1_eq_b1'].index(v)
    w = [0.0] * 8
    w[idx] = 1.0
    weights[f'l4.{v}.weight'] = torch.tensor([w], dtype=torch.float32)
    weights[f'l4.{v}.bias'] = torch.tensor([-0.5], dtype=torch.float32)

# Layer 5: For MIN, we want a_le_b = NOT a_gt_b
weights['l5.a_le_b.weight'] = torch.tensor([[-1.0, 1.0, 0.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l5.a_le_b.bias'] = torch.tensor([0.0], dtype=torch.float32)  # fires when a_gt_b=0 OR a_eq_b=1

for v in ['a1', 'a0', 'b1', 'b0']:
    idx = ['a_gt_b', 'a_eq_b', 'a1', 'a0', 'b1', 'b0'].index(v)
    w = [0.0] * 6
    w[idx] = 1.0
    weights[f'l5.{v}.weight'] = torch.tensor([w], dtype=torch.float32)
    weights[f'l5.{v}.bias'] = torch.tensor([-0.5], dtype=torch.float32)

# Layer 6: MUX - select a when a <= b, else b
# m1 = (a1 AND a_le_b) OR (b1 AND NOT a_le_b)
weights['l6.m1_a.weight'] = torch.tensor([[1.0, 1.0, 0.0, 0.0, 0.0]], dtype=torch.float32)
weights['l6.m1_a.bias'] = torch.tensor([-2.0], dtype=torch.float32)

weights['l6.m1_b.weight'] = torch.tensor([[-1.0, 0.0, 0.0, 1.0, 0.0]], dtype=torch.float32)
weights['l6.m1_b.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l6.m0_a.weight'] = torch.tensor([[1.0, 0.0, 1.0, 0.0, 0.0]], dtype=torch.float32)
weights['l6.m0_a.bias'] = torch.tensor([-2.0], dtype=torch.float32)

weights['l6.m0_b.weight'] = torch.tensor([[-1.0, 0.0, 0.0, 0.0, 1.0]], dtype=torch.float32)
weights['l6.m0_b.bias'] = torch.tensor([-1.0], dtype=torch.float32)

# Layer 7: Final OR
weights['l7.m1.weight'] = torch.tensor([[1.0, 1.0, 0.0, 0.0]], dtype=torch.float32)
weights['l7.m1.bias'] = torch.tensor([-1.0], dtype=torch.float32)

weights['l7.m0.weight'] = torch.tensor([[0.0, 0.0, 1.0, 1.0]], dtype=torch.float32)
weights['l7.m0.bias'] = torch.tensor([-1.0], dtype=torch.float32)

save_file(weights, 'model.safetensors')

# Verification
def min2(a1, a0, b1, b0):
    inp = torch.tensor([float(a1), float(a0), float(b1), float(b0)])

    l1_keys = ['a1_gt_b1', 'b1_gt_a1', 'a0_gt_b0', 'b0_gt_a0', 'both1_high', 'both1_low', 'a1', 'a0', 'b1', 'b0']
    l1 = {k: int((inp @ weights[f'l1.{k}.weight'].T + weights[f'l1.{k}.bias'] >= 0).item()) for k in l1_keys}
    l1_out = torch.tensor([float(l1[k]) for k in l1_keys])

    l2_keys = ['a1_eq_b1', 'a1_gt_b1', 'b1_gt_a1', 'a0_gt_b0', 'b0_gt_a0', 'a1', 'a0', 'b1', 'b0']
    l2 = {k: int((l1_out @ weights[f'l2.{k}.weight'].T + weights[f'l2.{k}.bias'] >= 0).item()) for k in l2_keys}
    l2_out = torch.tensor([float(l2[k]) for k in l2_keys])

    l3_keys = ['a_gt_b_part2', 'a0_neq_b0', 'a1_gt_b1', 'a1', 'a0', 'b1', 'b0', 'a1_eq_b1']
    l3 = {k: int((l2_out @ weights[f'l3.{k}.weight'].T + weights[f'l3.{k}.bias'] >= 0).item()) for k in l3_keys}
    l3_out = torch.tensor([float(l3[k]) for k in l3_keys])

    l4_keys = ['a_gt_b', 'a_eq_b', 'a1', 'a0', 'b1', 'b0']
    l4 = {k: int((l3_out @ weights[f'l4.{k}.weight'].T + weights[f'l4.{k}.bias'] >= 0).item()) for k in l4_keys}
    l4_out = torch.tensor([float(l4[k]) for k in l4_keys])

    l5_keys = ['a_le_b', 'a1', 'a0', 'b1', 'b0']
    l5 = {k: int((l4_out @ weights[f'l5.{k}.weight'].T + weights[f'l5.{k}.bias'] >= 0).item()) for k in l5_keys}
    l5_out = torch.tensor([float(l5[k]) for k in l5_keys])

    l6_keys = ['m1_a', 'm1_b', 'm0_a', 'm0_b']
    l6 = {k: int((l5_out @ weights[f'l6.{k}.weight'].T + weights[f'l6.{k}.bias'] >= 0).item()) for k in l6_keys}
    l6_out = torch.tensor([float(l6[k]) for k in l6_keys])

    m1 = int((l6_out @ weights['l7.m1.weight'].T + weights['l7.m1.bias'] >= 0).item())
    m0 = int((l6_out @ weights['l7.m0.weight'].T + weights['l7.m0.bias'] >= 0).item())

    return m1, m0

print("Verifying min2...")
errors = 0
for a in range(4):
    for b in range(4):
        a1, a0 = (a >> 1) & 1, a & 1
        b1, b0 = (b >> 1) & 1, b & 1
        m1, m0 = min2(a1, a0, b1, b0)
        result = 2*m1 + m0
        expected = min(a, b)
        if result != expected:
            errors += 1
            print(f"ERROR: min({a}, {b}) = {result}, expected {expected}")

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}")