# Conceptual Python code for linear symmetric quantization
import numpy as np

def quantize_tensor_symmetric(tensor_fp, num_bits):
    """
    Applies symmetric quantization to a floating-point tensor.
    Assumes zero_point = 0 for simplicity.
    """
    q_min = -(2**(num_bits - 1))
    q_max = (2**(num_bits - 1)) - 1

    # Determine scale factor
    abs_max_val = np.max(np.abs(tensor_fp))
    scale_factor = abs_max_val / q_max if q_max != 0 else 1.0

    # Quantize
    q_tensor = np.round(tensor_fp / scale_factor)

    # Clip to quantization range
    q_tensor = np.clip(q_tensor, q_min, q_max)

    return q_tensor.astype(np.int64), scale_factor, 0 # Returns quantized tensor, scale, zero_point

def dequantize_tensor_symmetric(q_tensor, scale_factor, zero_point):
    """
    Dequantizes a symmetric quantized tensor.
    """
    return (q_tensor - zero_point) * scale_factor

# Example usage
fp_weights = np.random.randn(4, 4) * 10
print(f"Original FP weights:\n{fp_weights}\n")

num_bits = 4
q_weights, scale, zero_point = quantize_tensor_symmetric(fp_weights, num_bits)
print(f"{num_bits}-bit Quantized weights:\n{q_weights}\n")
print(f"Scale factor: {scale}, Zero point: {zero_point}\n")

dequantized_weights = dequantize_tensor_symmetric(q_weights, scale, zero_point)
print(f"Dequantized FP weights:\n{dequantized_weights}\n")
print(f"Error (RMSE): {np.sqrt(np.mean((fp_weights - dequantized_weights)**2))}\n")