Code

(WIP) bitnet and t-mac#23540

Code

Open

liqunfu wants to merge 8 commits intomainmicrosoft/onnxruntime:mainfrom bitnetmicrosoft/onnxruntime:bitnetCopy head branch name to clipboard

Commits

Commits on Jan 30, 2025

• init code structure for matmul 2 bits

  Show description for 5484560
  liqunfu

  committed
  5484560

  Copy full SHA for 5484560

Commits on Jan 31, 2025

• add and pass q4dq tests for q2bit - rename file and test name later

  Show description for 8c1cfe1
  liqunfu

  committed
  8c1cfe1

  Copy full SHA for 8c1cfe1

• some fixes

  Show description for f6f22e3
  liqunfu

  committed
  f6f22e3

  Copy full SHA for f6f22e3

Commits on Feb 3, 2025

• add apis to neon and other avxs

  Show description for 3e1a951
  liqunfu

  committed
  3e1a951

  Copy full SHA for 3e1a951

• fix neon build

  Show description for 0130061
  liqunfu

  committed
  0130061

  Copy full SHA for 0130061

• disable 2bit test

  Show description for b4aad01
  liqunfu

  committed
  b4aad01

  Copy full SHA for b4aad01

Commits on Mar 7, 2025

• 2 bit quantize to support model builder

  Show description for ff531cb
  liqunfu

  committed
  ff531cb

  Copy full SHA for ff531cb

Commits on Mar 22, 2025

• mlas bitnet to use llama cpp dot routines. It serves as a fallback and shows how other bitnet implementations can be added to ORT. TODOs: handle any K values, blklen != 256, quantizer from fp32/16 …

  Show description for 34d4121
  liqunfu

  committed
  34d4121

  Copy full SHA for 34d4121