Code

Add Command R Plus support#6491

Code

Merged

ggerganov merged 15 commits intoggerganov:masterggerganov/llama.cpp:masterfrom RefractAI:masterRefractAI/llama.cpp:masterCopy head branch name to clipboardApr 9, 2024

+366-326

Commits

Commits on Apr 4, 2024

• Add Command R Plus GGUF

  

  S

  authored and

  S

  committed
  2efcd87

  Copy full SHA for 2efcd87

• Add Command R Plus GGUF

  

  S

  authored and

  S

  committed
  fbab984

  Copy full SHA for fbab984

• Loading works up to LayerNorm2D

  

  S

  authored and

  S

  committed
  e4b2e2d

  Copy full SHA for e4b2e2d

• Export new tensors in 1D so they are not quantized.

  

  S

  committed
  c354db7

  Copy full SHA for c354db7

Commits on Apr 5, 2024

• Fix embedding layer based on Noeda's example

  

  S

  committed
  553b09b

  Copy full SHA for 553b09b

• Whitespace

  

  S

  committed
  f3532ff

  Copy full SHA for f3532ff

• Add line

  

  S

  committed
  ec613b8

  Copy full SHA for ec613b8

Commits on Apr 6, 2024

• Fix unexpected tokens on MPS. Re-add F16 fix. ((Noeda)

  

  S

  committed
  c2658c3

  Copy full SHA for c2658c3

• dranger003: Fix block index overflow in CUDA dequantizing.

  

  S

  committed
  6745ea7

  Copy full SHA for 6745ea7

• Reverted blocked multiplication code as it still has issues and could affect other Llama arches

  

  S

  committed
  26e8f23

  Copy full SHA for 26e8f23

• export norms as f32

  slaren

  committed
  ce9413d

  Copy full SHA for ce9413d

• fix overflow issues during quant and other cleanup

  slaren

  committed
  78819c0

  Copy full SHA for 78819c0

• Merge pull request #1 from slaren/cmrp-fixes

  Show description for 8b6577b
  RefractAI

  authored
  8b6577b

  Copy full SHA for 8b6577b

Commits on Apr 7, 2024

• Type convention

  Show description for d292407
  

  RefractAI

  and

  ggerganov

  authored
  d292407

  Copy full SHA for d292407

Commits on Apr 8, 2024

• dranger003: Fix more int overflow during quant.

  

  S

  committed
  ea1aeba

  Copy full SHA for ea1aeba