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

@@ -32,7 +32,7 @@ services:
         - default
     container_name: hermes
     entrypoint: ["/bin/bash", "-c",
-      "bash /opt/data/hermes-tools/install.sh && . /opt/hermes/.venv/bin/activate && uv pip install openai 'mautrix[encryption]' -q && exec /usr/bin/tini -g -- /opt/hermes/docker/entrypoint.sh \"$@\"",
+      "bash /opt/data/hermes-tools/install.sh && exec /usr/bin/tini -g -- /opt/hermes/docker/entrypoint.sh \"$@\"",
       "hermes-entrypoint"]
     restart: always
     # Gateway run enables the internal API server on port 8642
@@ -58,8 +58,6 @@ services:
       - /mnt/HoardingCow_docker_data/Syncthing/telos-ro:/opt/data/telos-ro:ro
       # Syncthing-shared inbox — write tasks here, they sync to user's laptop
       - /mnt/HoardingCow_docker_data/Syncthing/telos-rw:/opt/data/telos-rw:rw
-      # Persist Python venv across container recreation (Matrix bridge deps, etc.)
-      - /mnt/HoardingCow_docker_data/Hermes/venv:/opt/hermes/.venv
     devices:
       - /dev/kfd:/dev/kfd
       - /dev/dri:/dev/dri
@@ -68,8 +66,6 @@ services:
       - "26"
     networks:
       - ai_backend
-    depends_on:
-      - honcho
 
   syncthing:
     image: syncthing/syncthing:latest
@@ -116,82 +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"
-
-  # --- Honcho: AI-native user modeling ---
-  honcho:
-    build: ./honcho
-    container_name: honcho
-    restart: unless-stopped
-    ports:
-      - "127.0.0.1:8000:8000"
-    environment:
-      - DB_CONNECTION_URI=postgresql+psycopg://honcho:honcho_pass@honcho-db:5432/honcho
-      - CACHE_URL=redis://honcho-redis:6379/0
-      - CACHE_ENABLED=true
-    volumes:
-      - /mnt/HoardingCow_docker_data/Honcho/data:/app/data
-    networks:
-      - ai_backend
-    depends_on:
-      honcho-db:
-        condition: service_healthy
-      honcho-redis:
-        condition: service_healthy
-
-  honcho-db:
-    image: pgvector/pgvector:pg15
-    container_name: honcho-db
-    restart: unless-stopped
-    ports:
-      - "127.0.0.1:5432:5432"
-    command: ["postgres", "-c", "max_connections=200"]
-    environment:
-      - POSTGRES_DB=honcho
-      - POSTGRES_USER=honcho
-      - POSTGRES_PASSWORD=honcho_pass
-      - PGDATA=/var/lib/postgresql/data/pgdata
-    volumes:
-      - /mnt/HoardingCow_docker_data/Honcho/postgres:/var/lib/postgresql/data
-      - ./honcho/init-db.sql:/docker-entrypoint-initdb.d/init.sql:ro
-    networks:
-      - ai_backend
-    healthcheck:
-      test: ["CMD-SHELL", "pg_isready -U honcho -d honcho"]
-      interval: 5s
-      timeout: 5s
-      retries: 5
-
-  honcho-redis:
-    image: redis:8
-    container_name: honcho-redis
-    restart: unless-stopped
-    ports:
-      - "127.0.0.1:6379:6379"
-    volumes:
-      - /mnt/HoardingCow_docker_data/Honcho/redis:/data
-    networks:
-      - ai_backend
-    healthcheck:
-      test: ["CMD-SHELL", "redis-cli ping"]
-      interval: 5s
-      timeout: 5s
-      retries: 5
 
 networks:
   ai_net:
@@ -201,47 +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/hermes/Dockerfile

@@ -48,14 +48,6 @@ RUN apt-get update && \
 # ---------- UV ----------
 COPY --chmod=0755 --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/
 
-WORKDIR /opt/hermes
-
-# ---------- Memory provider dependencies ----------
-# httpx: HTTP client for OpenViking plugin
-# honcho-ai: already installed in upstream image (v2.1.1+)
-RUN . /opt/hermes/.venv/bin/activate && \
-    uv pip install --no-cache-dir httpx
-
 # ---------- Piper TTS ----------
 RUN . /opt/hermes/.venv/bin/activate && \
     uv pip install --no-cache-dir piper-tts sounddevice numpy && \

ai/honcho/Dockerfile

@@ -1,42 +0,0 @@
-# build stage — fetches and builds Honcho from source
-# Using buildkit cache mounts for speed across rebuilds
-FROM python:3.13-slim-bookworm AS builder
-
-RUN apt-get update && \
-    apt-get install -y --no-install-recommends git && \
-    rm -rf /var/lib/apt/lists/*
-
-COPY --from=ghcr.io/astral-sh/uv:0.9.24 /uv /bin/uv
-
-# Clone Honcho at a pinned commit for reproducibility
-ARG HONCHO_REPO=https://github.com/plastic-labs/honcho
-ARG HONCHO_REF=main
-RUN git clone --depth 1 --branch ${HONCHO_REF} ${HONCHO_REPO} /app
-
-WORKDIR /app
-
-ENV UV_COMPILE_BYTECODE=1
-ENV UV_LINK_MODE=copy
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --no-group dev
-
-# --- runtime stage ---
-FROM python:3.13-slim-bookworm
-
-RUN groupadd --system app && \
-    useradd --system --gid app --create-home app
-
-COPY --from=builder /app /app
-COPY --from=builder /root/.cache/uv /root/.cache/uv
-
-WORKDIR /app
-ENV PATH="/app/.venv/bin:$PATH"
-ENV HOME=/app
-
-COPY --chown=app:app config.toml /app/config.toml
-
-USER app
-EXPOSE 8000
-
-CMD ["fastapi", "run", "--host", "0.0.0.0", "src/main.py"]

ai/honcho/config.toml

@@ -1,117 +0,0 @@
-[app]
-LOG_LEVEL = "INFO"
-MAX_MESSAGE_SIZE = 25000
-EMBED_MESSAGES = true
-NAMESPACE = "honcho"
-
-[db]
-CONNECTION_URI = "postgresql+psycopg://honcho:honcho_pass@honcho-db:5432/honcho"
-SCHEMA = "public"
-POOL_SIZE = 10
-MAX_OVERFLOW = 20
-
-[auth]
-USE_AUTH = false
-
-[sentry]
-ENABLED = false
-
-[telemetry]
-ENABLED = false
-
-[webhook]
-ENABLED = false
-
-[cache]
-ENABLED = true
-URL = "redis://honcho-redis:6379/0"
-
-[llm]
-DEFAULT_MAX_TOKENS = 4096
-
-# Embeddings via Ollama (nomic-embed-text recommended on this system)
-[embedding]
-VECTOR_DIMENSIONS = 768
-MAX_INPUT_TOKENS = 8192
-
-[embedding.model_config]
-transport = "openai"
-model = "nomic-embed-text"
-base_url = "http://ollama:11434/v1"
-
-# --- Deriver (user representation builder) ---
-[deriver]
-ENABLED = true
-WORKERS = 1
-POLLING_SLEEP_INTERVAL_SECONDS = 5.0
-FLUSH_ENABLED = true
-
-[deriver.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-# --- Dialectic ---
-[dialectic]
-MAX_INPUT_TOKENS = 4096
-SESSION_HISTORY_MAX_TOKENS = 8192
-
-[dialectic.levels.minimal]
-MAX_TOOL_ITERATIONS = 1
-MAX_OUTPUT_TOKENS = 512
-[dialectic.levels.minimal.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-[dialectic.levels.low]
-MAX_TOOL_ITERATIONS = 3
-[dialectic.levels.low.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-[dialectic.levels.medium]
-MAX_TOOL_ITERATIONS = 2
-[dialectic.levels.medium.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-[dialectic.levels.high]
-MAX_TOOL_ITERATIONS = 4
-[dialectic.levels.high.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-[dialectic.levels.max]
-MAX_TOOL_ITERATIONS = 10
-[dialectic.levels.max.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-# --- Summary ---
-[summary]
-ENABLED = true
-MESSAGES_PER_SHORT_SUMMARY = 20
-MESSAGES_PER_LONG_SUMMARY = 60
-
-[summary.model_config]
-transport = "openai"
-model = "hermes-3"
-base_url = "http://ollama:11434/v1"
-
-# --- Dream ---
-[dream]
-ENABLED = false
-
-# --- Peer Card ---
-[peer_card]
-ENABLED = true
-
-# --- Vector Store ---
-[vector_store]
-TYPE = "pgvector"
-DIMENSIONS = 768

ai/honcho/init-db.sql

@@ -1 +0,0 @@
-CREATE EXTENSION IF NOT EXISTS vector;

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"]