TL;DR: llama.cpp build b11715 replaces manual byte-permute logic with the __builtin_amdgcn_perm intrinsic when unpacking q1_0 blocks on the HIP/ROCm backend, a small but direct win for 1-bit quantized inference on AMD GPUs.
Summary: The ggml-org/llama.cpp release b11715 ships PR #29927, a ggml-cuda change that swaps in the AMDGPU __builtin_amdgcn_perm intrinsic for q1_0 unpacking in the HIP backend. The change targets the dequantization/data-layout path for 1-bit quantized weights, replacing generic bit-manipulation code with a native register-permute instruction. Alongside the source change, the release publishes standard prebuilt artifacts including an iOS XCFramework and UI bundle.
Why it matters: Indie builders running 1-bit (q1_0) models on AMD/ROCm hardware should see cheaper dequantization in the matmul pipeline, which matters most for memory-bound token generation. Rebuild or pull the b11715 artifacts and benchmark your q1_0 workloads against the prior build to confirm the gain on your GPU.
Source: github_releases