ggml/llama.cpp releases: b10171

TL;DR

A update to ggml/llama.cpp addresses an issue with Adreno KQ/KQV image kernels, which previously ignored dimension 3 entirely, by skipping these kernels for multi-stream batches. This fix is crucial for improving compatibility and performance on specific hardware configurations.

Detailed Summary

The ggml/llama.cpp project released an update that includes a fix for Adreno KQ/KQV image kernels, specifically addressing their handling of multi-stream batches by ignoring dimension 3 entirely. This update affects how sub-buffers are managed and kernel arguments are received. The broader impact is improved performance and functionality for GPU-accelerated operations on devices with Adreno GPUs.

Key Points

• • ggml/llama.cpp releases b10171
• • Skip Adreno KQ/KQV image kernels for multi-stream batches (#26189)
• • Adreno KQ/KQV image1d kernels ignore dim 3 entirely