[Question] How to reproduce QuIP# (No FT & No E_8)

[ChenMnZ]

Hi,

I wonder that how can I reproduce the results of QuIP#(no FT and no lattice codebook E8).

no lattice codebook E8 dose it mean uniform (INT) quantization, or is it still vector quantization.

[tsengalb99]

It means uniform scalar quantization. Specifically, for K bits, we round to the K bit grid [1/2, 3/2, … 2^K - 1/2] – 2^(K-1). You should be able to make a codebook class in lib/codebook that quantizes to this grid. You will need to determine the right scale. I have the code for this somewhere - ping me if you can’t implement a working codebook and I’ll try to find it. From: Mengzhao Chen ***@***.***> Sent: Monday, April 8, 2024 10:32 PM To: Cornell-RelaxML/quip-sharp ***@***.***> Cc: Subscribed ***@***.***> Subject: [Cornell-RelaxML/quip-sharp] [Question] How to reproduce QuIP# (No FT & No E_8) (Issue #52) Hi, I wonder that how can I reproduce the results of QuIP#(no FT and no lattice codebook E8). no lattice codebook E8 dose it mean uniform (INT) quantization, or is it still vector quantization. — Reply to this email directly, view it on GitHub <#52> , or unsubscribe <https://github.com/notifications/unsubscribe-auth/AH6WZSAVIIAK5LKFDO3Z663Y4NHJ3AVCNFSM6AAAAABF5XRCACVHI2DSMVQWIX3LMV43ASLTON2WKOZSGIZTENBZGAZDGMY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ChenMnZ]

Thanks for your quick reply, I'll try it.