Code

llama : initial Mamba-2 support#9126

Code

Open

compilade wants to merge 25 commits intomasterggml-org/llama.cpp:masterfrom compilade/mamba2ggml-org/llama.cpp:compilade/mamba2Copy head branch name to clipboard

+854-240

Commits

Commits on Aug 21, 2024

• llama : initial Mamba-2 support

  compilade

  committed
  1f0fea7

  Copy full SHA for 1f0fea7

• ggml : SIMD ggml_ssm_scan for Mamba-2

  Show description for dceff23
  compilade

  committed
  dceff23

  Copy full SHA for dceff23

• llama : support running Mamba-Codestral-7B-v0.1

  compilade

  committed
  2bfe9de

  Copy full SHA for 2bfe9de

• llama : fix Mamba-2 conv state saving

  Show description for aff9692
  compilade

  committed
  aff9692

  Copy full SHA for aff9692

Commits on Aug 22, 2024

• llama : remove unused variable

  compilade

  committed
  e04910d

  Copy full SHA for e04910d

• llama : add missing break

  compilade

  committed
  fa358e7

  Copy full SHA for fa358e7

• convert_hf : prefer SentencePiece tokenizer for Mamba-2 when present

  Show description for 38913dc
  compilade

  committed
  38913dc

  Copy full SHA for 38913dc

Commits on Sep 18, 2024

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  0e601ca

  Copy full SHA for 0e601ca

Commits on Oct 1, 2024

• llama : avoid redundant state copy for Mamba 1 and 2

  compilade

  committed
  273e7a4

  Copy full SHA for 273e7a4

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  7d6cb36

  Copy full SHA for 7d6cb36

Commits on Oct 2, 2024

• metal : attempt to adapt SSM_SCAN for Mamba-2

  compilade

  committed
  2c77d79

  Copy full SHA for 2c77d79

• metal : fix SSM_SCAN pipeline scope

  compilade

  committed
  87b97d0

  Copy full SHA for 87b97d0

• metal : use log and exp instead of log1pf and expf in SSM_SCAN

  compilade

  committed
  03d0e6e

  Copy full SHA for 03d0e6e

• metal : remove unused arguments for SSM_SCAN

  Show description for 7a351ab
  compilade

  committed
  7a351ab

  Copy full SHA for 7a351ab

• metal : add back n_seqs to SSM_SCAN args

  Show description for 8b15bc6
  compilade

  committed
  8b15bc6

  Copy full SHA for 8b15bc6

• metal : fix SSM_SCAN state head offset

  compilade

  committed
  5b8ec2b

  Copy full SHA for 5b8ec2b

Commits on Oct 3, 2024

• metal : fix wrong number of tokens per sequence in SSM_SCAN

  compilade

  committed
  62b09b3

  Copy full SHA for 62b09b3

Commits on Oct 12, 2024

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  038d958

  Copy full SHA for 038d958

• ggml : remove unused fast broadcast path in GGML_MUL

  Show description for 805512a
  compilade

  committed
  805512a

  Copy full SHA for 805512a

Commits on Nov 1, 2024

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  7d16e1b

  Copy full SHA for 7d16e1b

Commits on Nov 4, 2024

• ggml : avoid multiply by D in GGML_OP_SSM_SCAN

  Show description for 3bc7103
  compilade

  committed
  3bc7103

  Copy full SHA for 3bc7103

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  8d8f065

  Copy full SHA for 8d8f065

• convert : fix flake8 lint

  compilade

  committed
  b4e9c59

  Copy full SHA for b4e9c59

Commits on Nov 25, 2024

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  1ee6c48

  Copy full SHA for 1ee6c48

Commits on Feb 26, 2025

• Merge branch 'master' into compilade/mamba2

  compilade

  committed
  c9ecf62

  Copy full SHA for c9ecf62