RepoWatch / GitHub signal
llama.cpp b9553 release with RDNA3.5 HIP support
HIP RDNA3.5 support lands in latest llama.cpp.
Direct impact on local inference performance for AMD GPUs in our agent stack.
What changed
- New release tag b9553 (2026-06-07).
- Commit adding HIP support for gfx1152 and gfx1153 (RDNA3.5) on the default branch.
- Related: abetlen/llama-cpp-python v0.3.28-hip-radeon and abetlen/ggml-python v0.0.41-cu123 releases.
Why it matters
llama.cpp underpins local inference in OpenClaw, Hermes, and agent tooling. RDNA3.5 HIP additions improve AMD GPU compatibility for models like Gemma and Qwen.
My read
Clean release with targeted hardware support. The Python bindings updates align well. Not revolutionary but solid for our local stack.
Bottom line
Update now for AMD users. Worth testing in local environments before broader rollout.