ggml/llama.cpp releases: b10970

TL;DR

The ggml/llama.cpp project updated its codebase to use fp32 accumulators in fattn-mma on CDNA devices, enhancing performance. This update is significant for improving the efficiency of machine learning operations on specific hardware, making the software more versatile and potentially faster.

Detailed Summary

The ggml/llama.cpp project released a new update, implementing the use of fp32 accumulators in fattn-mma on CDNA devices, specifically to enhance performance. This update is part of commit b10970 and was merged through pull request #28576. The broader impact includes improved functionality for macOS Apple Silicon users, potentially enhancing computational efficiency in machine learning tasks.

Key Points

• • HIP: fattn-mma uses fp32 accumulation on MFMA devices
• • Use fp32 accumulators in fattn-mma on CDNA
• • macOS Apple Silicon (arm64) support available