v3.4.0

Based on ggml-org/llama.cpp b6735 (https://github.com/ggml-org/llama.cpp/releases/tag/b6735)

• Flash Attention for SWA models fixed
• New Flash Attention algorithm. It is optimized for long contexts (above 1024). See
  "Flash Attention algorithm selection" section for details how to select attention algorithm
  manually.

Also available at: DockerHub

v3.3.1

Also available at: DockerHub

v3.3.0

Also available at: DockerHub

v3.2.1

Also available at: DockerHub

v3.2.0

Also available at: DockerHub

v3.1.2

Also available at: DockerHub

v3.1.0

Also available at: DockerHub

v2.2.1

Update benchmark.py

v2.0.0

• Upgraded upstream tag enables Llama 3.1 in ollama
• Support for AmpereOne platform
• Breaking change: due to changed weight type IDs it is now required to re-quantize models to Q8R16 and Q4_K_4 formats with current llama-quantize tool.

v1.2.6

Create README.md