Skip to content

How is alternating optimization for MSE-optimized grids implemented in MR-GPTQ / FP-Quant? #25

@efsotr

Description

@efsotr

Hi, thanks for the great work.

I have a question about the implementation of Ingredient 1: MSE-Optimized Grids described in the MR-GPTQ paper.

In the paper, the objective is written as:

$$ \min_{s_T, s_{G_1},...,s_{G_k}} \sum_i \left\lVert \hat{X_i} - X_i \right\rVert_2^2 $$

The paper says:

We solve this by using alternating optimization over the block scales and the per-tensor scale, respectively.

Could you clarify how this alternating optimization is performed in practice?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions