ggml/llama.cpp releases: b12018

TL;DR

The ggml/llama.cpp project released version b12018, which includes a fix for f32 to q2_0/q4_0/q5_0 rounding in CPY; this update is important for improving the precision of quantization processes.

Detailed Summary

The ggml/llama.cpp project released a new version addressing rounding issues in CPY for f32 to q2_0/q4_0/q5_0 conversions. This update is part of the broader effort to enhance the library's performance on Apple Silicon (arm64) platforms, impacting users of macOS and iOS systems. The fix contributes to improving the accuracy and efficiency of quantization processes in the llamacpp framework.

Key Points

• • Fix implemented for f32 to q2_0/q4_0/q5_0 rounding in CPY (#31120)
• • Website: <https://llama.app>
• • Attestations available at: <https://github.com/ggml-org/llama.cpp/attestations/62012681>