create_safetensors.py

import torch
from safetensors.torch import save_file

weights = {}

# Minifloat Normalizer (4-bit mantissa)
# Shifts mantissa left until MSB is 1, adjusts exponent

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: M3,M2,M1,M0 (mantissa), E2,E1,E0 (exponent)
# Detect leading zeros
add_neuron('m3_is1', [1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], -1.0)  # M3=1, no shift
add_neuron('m2_lead', [-1.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], 0.0)  # M3=0,M2=1, shift 1
add_neuron('m1_lead', [-1.0, -1.0, 1.0, 0.0, 0.0, 0.0, 0.0], 1.0)  # shift 2
add_neuron('m0_lead', [-1.0, -1.0, -1.0, 1.0, 0.0, 0.0, 0.0], 2.0)  # shift 3

save_file(weights, 'model.safetensors')

def normalize(m3, m2, m1, m0, e2, e1, e0):
    m = m3*8 + m2*4 + m1*2 + m0
    e = e2*4 + e1*2 + e0

    if m == 0:
        return 0, 0, 0, 0, 0, 0, 0

    shift = 0
    while (m & 8) == 0 and shift < 4:
        m = (m << 1) & 0xF
        shift += 1

    e = max(0, e - shift)
    return (m>>3)&1, (m>>2)&1, (m>>1)&1, m&1, (e>>2)&1, (e>>1)&1, e&1

print("Verifying FP normalize...")
errors = 0
for m in range(16):
    for e in range(8):
        m3, m2, m1, m0 = (m>>3)&1, (m>>2)&1, (m>>1)&1, m&1
        e2, e1, e0 = (e>>2)&1, (e>>1)&1, e&1
        result = normalize(m3, m2, m1, m0, e2, e1, e0)
        # Verify MSB is 1 (or value is 0)
        if m != 0 and result[0] != 1:
            errors += 1

if errors == 0:
    print("All 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())}")