Metrics

qwantize.metrics.compute_metrics(W, W_dq, X=None)

Compute weight and output quantization error metrics.

Parameters:

• W – Original weights of shape (M, K).

• W_dq – Dequantized weights of shape (M, K).

• X – Input activations of shape (T, K). If provided, output error metrics are also computed.

Returns:

• "weight_error": ||W_dq - W||_F

• "weight_error_pct": ||W_dq - W||_F / ||W||_F * 100

• "output_error": ||W_dq @ X.T - W @ X.T||_F (only if X given)

• "output_error_pct": normalized output error in percent (only if X given)

Return type:

Dict with the following keys