Note Apply a transformation around the weight Q since this trans does not affects the nonlinearity it may be use to spare end incorporate into W matrix without adding errors.

Note To address these problems, we pro- pose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for var- ious quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end sup- port (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. https://github.com/usyd-fsalab/fp6_llm

Note https://github.com/MinorJerry/WebVoyager useful for automation see UI Path tipe of things.

Note Voxel power encoder decoder

Note Amazing one bit paper not quantisation but the model is build with 1-2 bit layers.

Note I think this is a paper that is worth reading because it is compatible with compactifAI they also have the code but it semas to be applied only on a toy model : https://arxiv.org/pdf/2402.17759.pdf

Note Sign-Value-Independent Decomposition I think this can be applied on top of the MPO

Note leader takes decision