fix: bump llama.cpp from b9596 to b9890 for slot/stability fixes

b9596 (Dec 2024) has known slot deadlock on first request causing
container freeze. b9890 includes:
- Slot selection consolidation (#24755)
- HIP fast-math support (#23862)
- Hundreds of stability and memory fixes

Compatible with ROCm 6.1 + gfx906 (same build flags).

ai/llama-cpp/Dockerfile

@@ -1,10 +1,12 @@
 # llama-cpp-rocm6/Dockerfile
 # Custom llama.cpp server with ROCm 6.1 + gfx906 (MI50) support.
 # Build: docker build -t llama-cpp:rocm-gfx906 .
+# NOTE: Update LLAMACPP_VERSION for bugfixes, but verify ROCm 6.1 compat
+#       by checking if recent releases still support -DGGML_HIP=ON.
 
 FROM rocm/dev-ubuntu-22.04:6.1.2-complete AS builder
 RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y curl git build-essential pkg-config cmake make && rm -rf /var/lib/apt/lists/*
-ARG LLAMACPP_VERSION=b9596
+ARG LLAMACPP_VERSION=b9890
 RUN git clone --depth 1 --branch ${LLAMACPP_VERSION} https://github.com/ggml-org/llama.cpp.git /build
 WORKDIR /build
 ENV HIP_PATH=/opt/rocm ROCM_PATH=/opt/rocm PATH=/opt/rocm/bin:/opt/rocm/llvm/bin:${PATH} CMAKE_PREFIX_PATH=/opt/rocm