ggml/llama.cpp releases: b10232

TL;DR

The ggml/llama.cpp project released a new version implementing DeepSeek V4 hyper-connections, optimizing kernels for better performance on Metal devices. This update is significant as it enhances computational efficiency in deep learning models, particularly beneficial for hardware that supports Metal.

Detailed Summary

The ggml/llama.cpp project has released a new version (b10232) that includes significant updates to DeepSeek V4 hyper-connections, specifically implementing optimized kernels for GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and GGML_OP_DSV4_HC_POST. This update also adds Metal dispatch support and tests the production Sinkhorn iteration process. These changes aim to enhance computational efficiency in deep learning models, particularly those utilizing hyper-connections.

Key Points

• • Implement GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and GGML_OP_DSV4_HC_POST with SIMDgroup register and shuffle optimized kernels.
• • Add Metal dispatch and support plumbing for DeepSeek V4 hyper-connections.
• • Test the production Sinkhorn iteration coupling.