GPTQ-2D: Cubic-Time Two-Sided Adaptive Rounding
GPTQ-2D matches the vectorized two-sided rounding result while cutting runtime from quartic to cubic.
The paper studies adaptive rounding for matrices under a two-sided quadratic metric, where fixed nonsingular bases act on both sides of the residual. A direct vectorized approach can use the standard one-dimensional algorithm, but the Kronecker-product metric makes it quartic in the matrix dimension. GPTQ-2D instead rounds entries anti-diagonal by anti-diagonal, with entries on each anti-diagonal independent enough to round in parallel. Source: HF Daily Papers' note.
The paper studies adaptive rounding for matrices under a two-sided quadratic metric, where fixed nonsingular bases act on both sides of the residual. A direct vectorized approach can use the standard one-dimensional algorithm, but the Kronecker-product metric makes it quartic in the matrix dimension. GPTQ-2D instead rounds entries anti-diagonal by anti-diagonal, with entries on each anti-diagonal independent enough to round in parallel. Source: HF Daily Papers' note.
score 4