b8911

Llama.cpp's latest release, b8911, brings significant enhancements to Hexagon processor support, crucial for running LLMs on Qualcomm Snapdragon-powered devices. The update introduces comprehensive profiling capabilities for Hexagon DSPs and NPUs, allowing developers to measure and optimize performance at a granular level. Key changes include a fully asynchronous profiler that replaces the older opsync approach, configurable Performance Monitoring Unit (PMU) events, and a new post-processing Python tool (`ggml-hexagon-profile.py`) for analyzing profile logs.

The release refactors op-batch queue handling and introduces separate profile descriptors for basic and extended profiling modes. Notably, the profiler can now be enabled from the host via interface hooks, and PMU counter collection is configurable to avoid performance overhead on older devices. The update also simplifies output formatting and adds environment variable support for setting PMU events. This work, contributed by developers including Sigbjørn Skjæret and Trivikram Reddy, marks a significant step in making LLM inference efficient on mobile and edge devices.