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/Dockerfile

@@ -1,68 +0,0 @@
-FROM ghcr.io/astral-sh/uv:0.11.6-python3.13-trixie@sha256:b3c543b6c4f23a5f2df22866bd7857e5d304b67a564f4feab6ac22044dde719b AS uv_source
-FROM tianon/gosu:1.19-trixie@sha256:3b176695959c71e123eb390d427efc665eeb561b1540e82679c15e992006b8b9 AS gosu_source
-FROM debian:13.4
-
-# Disable Python stdout buffering to ensure logs are printed immediately
-ENV PYTHONUNBUFFERED=1
-
-# Store Playwright browsers outside the volume mount so the build-time
-# install survives the /opt/data volume overlay at runtime.
-ENV PLAYWRIGHT_BROWSERS_PATH=/opt/hermes/.playwright
-
-# Install system dependencies in one layer, clear APT cache
-# tini reaps orphaned zombie processes (MCP stdio subprocesses, git, bun, etc.)
-# that would otherwise accumulate when hermes runs as PID 1. See #15012.
-RUN apt-get update && \
-    apt-get install -y --no-install-recommends \
-        build-essential nodejs npm python3 ripgrep ffmpeg gcc python3-dev libffi-dev procps git openssh-client docker-cli tini \
-        curl poppler-utils imagemagick emacs-nox qemu-user-static binfmt-support qemu-user-binfmt && \
-        texlive-latex-base texlive-latex-extra texlive-fonts-recommended texlive-xetex texlive-science \
-        chromium xvfb \
-        fonts-noto-color-emoji fonts-unifont fonts-liberation fonts-ipafont-gothic fonts-wqy-zenhei fonts-tlwg-loma-otf fonts-freefont-ttf \
-        libasound2t64 libatk-bridge2.0-0t64 libatk1.0-0t64 libatspi2.0-0t64 libcairo2 libcups2t64 libdbus-1-3 libdrm2 libgbm1 libglib2.0-0t64 libnspr4 libnss3 libpango-1.0-0 libx11-6 libxcb1 libxcomposite1 libxdamage1 libxext6 libxfixes3 libxkbcommon0 libxrandr2 && \
-    rm -rf /var/lib/apt/lists/*
-
-# Non-root user for runtime; UID can be overridden via HERMES_UID at runtime
-RUN useradd -u 10000 -m -d /opt/data hermes
-
-COPY --chmod=0755 --from=gosu_source /gosu /usr/local/bin/
-COPY --chmod=0755 --from=uv_source /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/
-
-WORKDIR /opt/hermes
-
-# ---------- Layer-cached dependency install ----------
-# Copy only package manifests first so npm install + Playwright are cached
-# unless the lockfiles themselves change.
-COPY package.json package-lock.json ./
-COPY web/package.json web/package-lock.json web/
-
-RUN npm install --prefer-offline --no-audit && \
-    npx playwright install --with-deps chromium --only-shell && \
-    (cd web && npm install --prefer-offline --no-audit) && \
-    npm cache clean --force
-
-# ---------- Source code ----------
-# .dockerignore excludes node_modules, so the installs above survive.
-COPY --chown=hermes:hermes . .
-
-# Build web dashboard (Vite outputs to hermes_cli/web_dist/)
-RUN cd web && npm run build
-
-# ---------- Permissions ----------
-# Make install dir world-readable so any HERMES_UID can read it at runtime.
-# The venv needs to be traversable too.
-USER root
-RUN chmod -R a+rX /opt/hermes
-# Start as root so the entrypoint can usermod/groupmod + gosu.
-# If HERMES_UID is unset, the entrypoint drops to the default hermes user (10000).
-
-# ---------- Python virtualenv ----------
-RUN uv venv && \
-    uv pip install --no-cache-dir -e ".[all]"
-
-# ---------- Runtime ----------
-ENV HERMES_WEB_DIST=/opt/hermes/hermes_cli/web_dist
-ENV HERMES_HOME=/opt/data
-ENV PATH="/opt/data/.local/bin:${PATH}"
-VOLUME [ "/opt/data" ]
-ENTRYPOINT [ "/usr/bin/tini", "-g", "--", "/opt/hermes/docker/entrypoint.sh" ]

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:
+  llama-cpp-hermes:
-  #   image: ghcr.io/ggml-org/llama.cpp:server-rocm
+    image: llama-cpp:rocm-gfx906
-  #   container_name: llama_cpp_devstral
+    container_name: llama-cpp-hermes
-  #   restart: unless-stopped
+    restart: unless-stopped
-  #   networks:
+    networks:
-  #     - ai_backend
+      - ai_backend
-  #   ports:
+    ports:
-  #     - "8300:8080"
+      - "127.0.0.1:8300:8080"
-  #   ipc: host
+    ipc: host
-  #   devices:
+    devices:
-  #     - "/dev/kfd:/dev/kfd"
+      - /dev/kfd:/dev/kfd
-  #     - "/dev/dri:/dev/dri"
+      - /dev/dri:/dev/dri
-  #   group_add:
+    group_add:
-  #     - "303" # video
+      - "303"
-  #     - "26"  # render
+      - "26"
-  #   environment:
+    environment:
-  #     HSA_OVERRIDE_GFX_VERSION: 9.0.6
+      - HSA_OVERRIDE_GFX_VERSION=9.0.6
-  #     HIP_VISIBLE_DEVICES: 0,1
+      - HSA_ENABLE_SDMA=0
-  #     LLAMA_CACHE: /models
+      - HIP_VISIBLE_DEVICES=0,1
-  #   volumes:
+      - LLAMA_CACHE=/models
-  #     - /mnt/HoardingCow_docker_data/Llama_cpp/models:/models
+    volumes:
-  #     - /mnt/HoardingCow_docker_data/Llama_cpp/devstral-agent.jinja:/template.jinja
+      - /mnt/HoardingCow_docker_data/Llama_cpp/models:/models
-  #   command: >
+      - /mnt/HoardingCow_docker_data/Ollama/ollama/models/blobs/sha256-17823599694fa3503ef54bf748d5078c6ce881f4d01616cafa255dc05d215a08:/model.gguf:ro
-  #     -hf unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Devstral-Small-2-24B-Instruct-2512-Q8_0.gguf
+    command: >
-  #     -a devstral-2-small-llama_cpp
+      -m /model.gguf
-  #     --chat-template-file /template.jinja
+      --host 0.0.0.0
-  #     --host 0.0.0.0
+      --port 8080
-  #     --port 8080
+      --gpu-layers 99
-  #     --n-gpu-layers 99
+      --ctx-size 163840
-  #     --ctx-size 163840
+      -ctk f16 -ctv f16
-  #     --batch-size 4096
+      --flash-attn on
-  #     --ubatch-size 4096
+      --split-mode layer
-  #     --cache-type-k f16
+      --no-mmap
-  #     --cache-type-v f16
+      --n-predict -1
-  #     --cache-reuse 256
-  #     --flash-attn on
-  #     --context-shift
-  #     --split-mode layer
-  #     --no-mmap
-  #     --n-predict -1
-  #     --parallel 2
 
   # vllm:
   #   image: nalanzeyu/vllm-gfx906:v0.9.0-rocm6.3

ai/hermes/entrypoint.sh

@@ -1,20 +0,0 @@
-#!/bin/bash
-set -e
-
-# Hermes Agent entrypoint script
-# Installs custom tools and runtime dependencies,
-# then delegates to the passed command (usually "gateway run").
-
-# Install custom tools from persistent volume if available
-if [ -f /opt/data/hermes-tools/install.sh ]; then
-    bash /opt/data/hermes-tools/install.sh
-fi
-
-# Install additional runtime deps (idempotent)
-if command -v uv &>/dev/null; then
-    uv pip install --system --no-cache-dir --quiet \
-        openai mautrix[encryption] 2>/dev/null || true
-fi
-
-# Execute the passed command with tini for proper signal handling
-exec tini -g -- "$@"

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"]