feat: add llama-cpp-hermes service with ROCm 6.1 + gfx906 support

- Add custom llama.cpp Dockerfile with ROCm 6.1 + gfx906 (MI50) build
- Add llama-cpp-hermes service serving Hermes 4.3 on dual MI50 GPUs
- Strip GPU devices/ROCm env from ollama service (CPU-only for embeddings)

Hermes 4.3 runs at ~19 t/s on dual MI50s with 160K context.

ai/compose.yml

@@ -112,22 +112,7 @@ services:
       - /mnt/HoardingCow_docker_data/Ollama/ollama:/root/.ollama
     environment:
       - OLLAMA_VULKAN=0
-      - HSA_OVERRIDE_GFX_VERSION=9.0.6
-      - HCC_AMDGPU_TARGET=gfx906
-      - HIP_VISIBLE_DEVICES=0,1
-      - ROCR_VISIBLE_DEVICES=0,1
-      - HSA_ENABLE_SDMA=0 
       - OLLAMA_HOST=0.0.0.0
-      - OLLAMA_DEBUG=1
-      - OLLAMA_FLASH_ATTENTION=1
-      - OLLAMA_NUM_PARALLEL=2
-    devices:
-      # Map the render nodes and KFD for ROCm to work inside the container
-      - /dev/kfd:/dev/kfd
-      - /dev/dri:/dev/dri
-    group_add:
-      - "303"
-      - "26"
 
 networks:
   ai_net:
@@ -137,47 +122,40 @@ networks:
     driver: bridge
     name: ai_backend
     
-  # llama_cpp_devstral:
-  #   image: ghcr.io/ggml-org/llama.cpp:server-rocm
-  #   container_name: llama_cpp_devstral
-  #   restart: unless-stopped
-  #   networks:
-  #     - ai_backend
-  #   ports:
-  #     - "8300:8080"
-  #   ipc: host
-  #   devices:
-  #     - "/dev/kfd:/dev/kfd"
-  #     - "/dev/dri:/dev/dri"
-  #   group_add:
-  #     - "303" # video
-  #     - "26"  # render
-  #   environment:
-  #     HSA_OVERRIDE_GFX_VERSION: 9.0.6
-  #     HIP_VISIBLE_DEVICES: 0,1
-  #     LLAMA_CACHE: /models
-  #   volumes:
-  #     - /mnt/HoardingCow_docker_data/Llama_cpp/models:/models
-  #     - /mnt/HoardingCow_docker_data/Llama_cpp/devstral-agent.jinja:/template.jinja
-  #   command: >
-  #     -hf unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Devstral-Small-2-24B-Instruct-2512-Q8_0.gguf
-  #     -a devstral-2-small-llama_cpp
-  #     --chat-template-file /template.jinja
-  #     --host 0.0.0.0
-  #     --port 8080
-  #     --n-gpu-layers 99
-  #     --ctx-size 163840
-  #     --batch-size 4096
-  #     --ubatch-size 4096
-  #     --cache-type-k f16
-  #     --cache-type-v f16
-  #     --cache-reuse 256
-  #     --flash-attn on
-  #     --context-shift
-  #     --split-mode layer
-  #     --no-mmap
-  #     --n-predict -1
-  #     --parallel 2
+  llama-cpp-hermes:
+    image: llama-cpp:rocm-gfx906
+    container_name: llama-cpp-hermes
+    restart: unless-stopped
+    networks:
+      - ai_backend
+    ports:
+      - "127.0.0.1:8300:8080"
+    ipc: host
+    devices:
+      - /dev/kfd:/dev/kfd
+      - /dev/dri:/dev/dri
+    group_add:
+      - "303"
+      - "26"
+    environment:
+      - HSA_OVERRIDE_GFX_VERSION=9.0.6
+      - HSA_ENABLE_SDMA=0
+      - HIP_VISIBLE_DEVICES=0,1
+      - LLAMA_CACHE=/models
+    volumes:
+      - /mnt/HoardingCow_docker_data/Llama_cpp/models:/models
+      - /mnt/HoardingCow_docker_data/Ollama/ollama/models/blobs/sha256-17823599694fa3503ef54bf748d5078c6ce881f4d01616cafa255dc05d215a08:/model.gguf:ro
+    command: >
+      -m /model.gguf
+      --host 0.0.0.0
+      --port 8080
+      --gpu-layers 99
+      --ctx-size 163840
+      -ctk f16 -ctv f16
+      --flash-attn on
+      --split-mode layer
+      --no-mmap
+      --n-predict -1
 
   # vllm:
   #   image: nalanzeyu/vllm-gfx906:v0.9.0-rocm6.3

ai/llama-cpp/Dockerfile

@@ -0,0 +1,30 @@
+# llama-cpp-rocm6/Dockerfile
+# Custom llama.cpp server with ROCm 6.1 + gfx906 (MI50) support.
+# Build: docker build -t llama-cpp:rocm-gfx906 .
+
+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
+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
+RUN mkdir build && cd build && \
+    cmake .. -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release \
+      -DAMDGPU_TARGETS="gfx906:xnack-" \
+      -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+      -DGGML_CUDA=OFF -DGGML_VULKAN=OFF -DGGML_METAL=OFF \
+      -DBUILD_SHARED_LIBS=OFF && \
+    cmake --build . --target llama-server -- -j $(nproc)
+
+FROM ubuntu:24.04
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y \
+    ca-certificates curl libstdc++6 libgomp1 libopenblas0 \
+    libnuma1 libelf1 libdrm2 libdrm-amdgpu1 \
+    && rm -rf /var/lib/apt/lists/*
+COPY --from=builder /opt/rocm/lib/ /opt/rocm/lib/
+COPY --from=builder /opt/rocm/share/ /opt/rocm/share/
+COPY --from=builder /build/build/bin/llama-server /usr/local/bin/llama-server
+RUN echo /opt/rocm/lib > /etc/ld.so.conf.d/rocm.conf && ldconfig
+ENV HSA_OVERRIDE_GFX_VERSION=9.0.6 HCC_AMDGPU_TARGET=gfx906 HSA_ENABLE_SDMA=0
+EXPOSE 8080
+ENTRYPOINT ["/usr/local/bin/llama-server"]