ggml/llama.cpp releases: b11866

TL;DR

A recent update to ggml/llama.cpp removed the need for a specific rescaling method, optimizing memory efficiency in the model's computation process. This change is significant as it enhances the performance and resource management of large language models.

Detailed Summary

The ggml/llama.cpp project released a new version (b11866) that removes the `mul_mm_id` amax rescaling in favor of per-expert-token amax scaling, improving MoE down-projection activations. This update involves contributions from metal and others on GitHub. The change is expected to enhance the efficiency and performance of large language models implemented with ggml/llama.cpp.

Key Points

• • metal: remove mul_mm_id amax rescaling
• • src1 power-of-two rescale superseded by per expert-token amax scaling
• • MoE down-projection activations improved