halo-box/strix-llama.cpp

Shipped this two weeks ago already, but never shared it. Both Halo-box and the Gyro llama.cpp fork have a dashboard to see performance and request stats.

strix-llama new branch: prompt processing +21–54% over Vulkan and up to 2.7× over stock ROCm kernel at 128k, same outputs quality

strix-llama new branch: prompt processing +21–54% over Vulkan and up to 2.7× over stock ROCm at 128k, same outputs

For Strix Halo - Official llama.cpp isn't ideal and how to highest possible throughput

I made PP 40% faster for MoE and 20% faster for dense models