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

@@ -44,7 +44,7 @@ services:
       - API_SERVER_HOST=0.0.0.0
       - API_SERVER_KEY=hermes_local_key
       - GATEWAY_ALLOW_ALL_USERS=true
-      - OPENROUTER_API_KEY=${OPEN...KEY}
+      - OPENROUTER_API_KEY=${OPENROUTER_API_KEY}
       # ROCm for GPU-accelerated faster-whisper STT
       - HSA_OVERRIDE_GFX_VERSION=9.0.6
       - HCC_AMDGPU_TARGET=gfx906
@@ -112,166 +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"
-
-  paperclip-db:
-    image: postgres:17-alpine
-    container_name: paperclip-db
-    restart: always
-    environment:
-      POSTGRES_USER: paperclip
-      POSTGRES_PASSWORD: ${PAPERCLIP_DB_PASSWORD:?PAPERCLIP_DB_PASSWORD must be set}
-      POSTGRES_DB: paperclip
-    healthcheck:
-      test: ["CMD-SHELL", "pg_isready -U paperclip -d paperclip"]
-      interval: 5s
-      timeout: 5s
-      retries: 10
-    volumes:
-      - /mnt/HoardingCow_docker_data/Paperclip/pgdata:/var/lib/postgresql/data
-    networks:
-      - ai_backend
-
-  paperclip:
-    image: ghcr.io/paperclipai/paperclip:v2026.517.0
-    container_name: paperclip
-    restart: always
-    ports:
-      - "127.0.0.1:3100:3100"
-    environment:
-      - HOST=0.0.0.0
-      - PORT=3100
-      - SERVE_UI=true
-      - DATABASE_URL=postgres://paperclip:***@paperclip-db:5432/paperclip
-      - BETTER_AUTH_SECRET=${PAPE...CRET must be set}
-      - PAPERCLIP_PUBLIC_URL=https://paperclip.lazyworkhorse.net
-      - PAPERCLIP_DEPLOYMENT_MODE=authenticated
-      - PAPERCLIP_DEPLOYMENT_EXPOSURE=private
-    volumes:
-      - /mnt/HoardingCow_docker_data/Paperclip/data:/paperclip
-    depends_on:
-      paperclip-db:
-        condition: service_healthy
-    networks:
-      - ai_net
-      - ai_backend
-    labels:
-      - "traefik.enable=true"
-      - "traefik.docker.network=ai_net"
-
-      - "traefik.http.routers.paperclip-http.rule=Host(`paperclip.lazyworkhorse.net`)"
-      - "traefik.http.routers.paperclip-http.entrypoints=web"
-      - "traefik.http.routers.paperclip-http.middlewares=redirect-to-https"
-
-      - "traefik.http.routers.paperclip-https.rule=Host(`paperclip.lazyworkhorse.net`)"
-      - "traefik.http.routers.paperclip-https.entrypoints=websecure"
-      - "traefik.http.routers.paperclip-https.tls=true"
-      - "traefik.http.routers.paperclip-https.tls.certresolver=njalla"
-
-      - "traefik.http.services.paperclip.loadbalancer.server.port=3100"
-
-  # ---------------------------------------------------------------------------
-  # Honcho — Memory infrastructure for stateful AI agents
-  # Self-hosted memory server with pgvector for embedding storage.
-  # Defaults to Ollama for embeddings; configure LLM provider for full deriver
-  # and summarization support.
-  #
-  # API port: 8000
-  # Web:       https://honcho.lazyworkhorse.net
-  # Docs:      https://github.com/plastic-labs/honcho
-  # ---------------------------------------------------------------------------
-
-  honcho-db:
-    image: pgvector/pgvector:pg17-trixie
-    container_name: honcho-db
-    restart: unless-stopped
-    environment:
-      POSTGRES_DB: honcho
-      POSTGRES_USER: honcho
-      POSTGRES_PASSWORD: ${HONCHO_DB_PASSWORD:?HONCHO_DB_PASSWORD must be set}
-    healthcheck:
-      test: ["CMD-SHELL", "pg_isready -U honcho -d honcho"]
-      interval: 5s
-      timeout: 5s
-      retries: 10
-    volumes:
-      - /mnt/HoardingCow_docker_data/Honcho/pgdata:/var/lib/postgresql/data
-      - ./honcho/init.sql:/docker-entrypoint-initdb.d/init.sql
-    networks:
-      - ai_backend
-
-  honcho:
-    build:
-      context: ./honcho
-      dockerfile: Dockerfile
-    container_name: honcho
-    restart: unless-stopped
-    ports:
-      - "127.0.0.1:8000:8000"
-    depends_on:
-      honcho-db:
-        condition: service_healthy
-    environment:
-      DB_CONNECTION_URI: postgresql+psycopg://honcho:${HONCHO_DB_PASSWORD:?HONCHO_DB_PASSWORD must be set}@honcho-db:5432/honcho
-      LOG_LEVEL: INFO
-      LLM_OPENAI_API_KEY: ${LLM_OPENAI_API_KEY:-ollama}
-    volumes:
-      - /mnt/HoardingCow_docker_data/Honcho/config.toml:/app/config.toml
-    networks:
-      - ai_backend
-      - ai_net
-    labels:
-      - "traefik.enable=true"
-      - "traefik.docker.network=ai_net"
-
-      - "traefik.http.routers.honcho-http.rule=Host(`honcho.lazyworkhorse.net`)"
-      - "traefik.http.routers.honcho-http.entrypoints=web"
-      - "traefik.http.routers.honcho-http.middlewares=redirect-to-https"
-
-      - "traefik.http.routers.honcho-https.rule=Host(`honcho.lazyworkhorse.net`)"
-      - "traefik.http.routers.honcho-https.entrypoints=websecure"
-      - "traefik.http.routers.honcho-https.tls=true"
-      - "traefik.http.routers.honcho-https.tls.certresolver=njalla"
-
-      - "traefik.http.services.honcho.loadbalancer.server.port=8000"
-
-  holographic-memory:
-    build:
-      context: ./holographic-memory
-    image: holographic-memory:latest
-    container_name: holographic-memory
-    restart: unless-stopped
-    ports:
-      - "127.0.0.1:8100:8100"
-    environment:
-      - HOLOGRAPHIC_DB_PATH=/data/holographic/memory_store.db
-      - HOLOGRAPHIC_PORT=8100
-      - HOLOGRAPHIC_DEFAULT_TRUST=0.5
-    volumes:
-      - /mnt/HoardingCow_docker_data/HolographicMemory:/data/holographic
-    networks:
-      - ai_backend
-    healthcheck:
-      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8100/health')"]
-      interval: 30s
-      timeout: 5s
-      retries: 3
-      start_period: 10s
 
 networks:
   ai_net:
@@ -281,47 +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
@@ -424,8 +258,8 @@ networks:
   #     - /home/gortium/infra:/data/workspace/infra
   #   environment:
   #     - TZ=America/Toronto
-  #     - OPENCLAW_GATEWAY_TOKEN=${OPEN...KEN}
+  #     - OPENCLAW_GATEWAY_TOKEN=${OPENCLAW_GATEWAY_TOKEN}
-  #     - OPENROUTER_API_KEY=${OPEN...KEY}
+  #     - OPENROUTER_API_KEY=${OPENROUTER_API_KEY}
   #     # Point to the sidecar browser
   #     - BROWSER_CDP_URL=http://openclaw-browser:9222
   #     - BROWSER_EVALUATE_ENABLED=true
@@ -470,7 +304,7 @@ networks:
   #     - PGID=1000
   #     - PUBLIC_KEY_FILE=/config/ssh/authorized_keys
   #     - SUDO_ACCESS=false
-  #     - PASSWORD_ACCESS=***
+  #     - PASSWORD_ACCESS=false
   #   volumes:
   #     - /mnt/HoardingCow_docker_data/openclaw/ssh-config:/config
   #     - /home/gortium/infra:/data/workspace/infra:ro

ai/hermes/Dockerfile

@@ -79,26 +79,6 @@ PYEOF
 COPY --chmod=0755 himalaya-ro.sh /usr/local/bin/himalaya-ro
 
 
-# ---------- Install 7-Zip (7zz) for CHM extraction ----------
-RUN /opt/hermes/.venv/bin/python3 /dev/stdin << 'PYEOF'
-import urllib.request, tarfile, os, shutil, subprocess
-url = 'https://github.com/ip7z/7zip/releases/download/26.01/7z2601-linux-x64.tar.xz'
-xz = '/tmp/7z2601-linux-x64.tar.xz'
-urllib.request.urlretrieve(url, xz)
-os.makedirs('/tmp/7z', exist_ok=True)
-with tarfile.open(xz, 'r:xz') as t:
-    t.extractall('/tmp/7z')
-shutil.move('/tmp/7z/7zz', '/usr/local/bin/7zz')
-os.chmod('/usr/local/bin/7zz', 0o755)
-shutil.rmtree('/tmp/7z', ignore_errors=True)
-os.remove(xz)
-# Verify
-result = subprocess.run(['/usr/local/bin/7zz'], capture_output=True, text=True)
-assert result.returncode == 0, f'7zz verify failed: {result.stderr}'
-print('7-Zip 26.01 installed successfully')
-PYEOF
-
-
 # ---------- Runtime ----------
 USER hermes
 ENV HERMES_HOME=/opt/data

ai/honcho/Dockerfile

@@ -1,72 +0,0 @@
-# Honcho — Memory infrastructure for stateful AI agents
-# Builds the Honcho FastAPI server from the official GitHub repository.
-#
-# Usage:
-#   docker compose build honcho
-#   docker compose up honcho
-#
-# Reference: https://github.com/plastic-labs/honcho
-
-# ---------------------------------------------------------------------------
-# Stage 1 — clone source & install dependencies
-# ---------------------------------------------------------------------------
-FROM python:3.13-slim-bookworm AS builder
-
-RUN apt-get update && apt-get install -y git && rm -rf /var/lib/apt/lists/*
-
-COPY --from=ghcr.io/astral-sh/uv:0.9.24 /uv /bin/uv
-
-WORKDIR /src
-RUN git clone --depth 1 --branch main https://github.com/plastic-labs/honcho.git .
-
-ENV UV_COMPILE_BYTECODE=1
-ENV UV_LINK_MODE=copy
-
-# Install project dependencies (frozen from lockfile, no dev)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --no-install-project --no-group dev
-
-# ---------------------------------------------------------------------------
-# Stage 2 — runtime image
-# ---------------------------------------------------------------------------
-FROM python:3.13-slim-bookworm AS runtime
-
-COPY --from=ghcr.io/astral-sh/uv:0.9.24 /uv /bin/uv
-
-RUN apt-get update && \
-    apt-get install -y --no-install-recommends ca-certificates && \
-    rm -rf /var/lib/apt/lists/*
-
-WORKDIR /app
-
-ENV UV_COMPILE_BYTECODE=1
-ENV UV_LINK_MODE=copy
-ENV PYTHONDONTWRITEBYTECODE=1
-ENV PYTHONUNBUFFERED=1
-ENV PATH="/app/.venv/bin:$PATH"
-ENV HOME=/app
-ENV UV_CACHE_DIR=/tmp/uv-cache
-
-# Copy the dependency layer from the builder
-COPY --from=builder /src/uv.lock /src/pyproject.toml /app/
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --no-group dev
-
-# Copy application source and config
-COPY --from=builder /src/src/ /app/src/
-COPY --from=builder /src/migrations/ /app/migrations/
-COPY --from=builder /src/scripts/ /app/scripts/
-COPY --from=builder /src/docker/ /app/docker/
-COPY --from=builder /src/alembic.ini /app/alembic.ini
-
-# Create non-root user
-RUN addgroup --system app && \
-    adduser --system --ingroup app app && \
-    mkdir -p /tmp/uv-cache && \
-    chown -R app:app /app /tmp/uv-cache
-
-USER app
-EXPOSE 8000
-
-# The entrypoint.sh script runs database migrations then starts the FastAPI server
-ENTRYPOINT ["sh", "docker/entrypoint.sh"]

ai/honcho/init.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"]

build/honcho/config.toml

@@ -1,93 +0,0 @@
-# Honcho Configuration
-# Pre-configured for self-hosted deployment with Ollama embeddings.
-# Mount this file at /app/config.toml in the Honcho container.
-#
-# Environment variables override these values at runtime
-# (e.g. DB_CONNECTION_URI, DERIVER_*).
-
-[app]
-LOG_LEVEL = "INFO"
-NAMESPACE = "honcho"
-SESSION_OBSERVERS_LIMIT = 10
-GET_CONTEXT_MAX_TOKENS = 16384
-EMBED_MESSAGES = true
-
-[db]
-# Connection URI is set via environment variable DB_CONNECTION_URI
-SCHEMA = "public"
-POOL_SIZE = 10
-MAX_OVERFLOW = 20
-POOL_TIMEOUT = 30
-POOL_RECYCLE = 300
-POOL_PRE_PING = true
-POOL_USE_LIFO = true
-SQL_DEBUG = false
-
-[auth]
-USE_AUTH = false
-
-[llm]
-DEFAULT_MAX_TOKENS = 4096
-
-[embedding]
-VECTOR_DIMENSIONS = 768
-MAX_INPUT_TOKENS = 8192
-MAX_TOKENS_PER_REQUEST = 2048
-
-[embedding.model_config]
-transport = "openai"
-model = "nomic-embed-text:latest"
-
-[embedding.model_config.overrides]
-base_url = "http://ollama:11434/v1"
-# Ollama does not require an API key; env var must be set to non-empty string
-api_key_env = "LLM_OPENAI_API_KEY"
-
-[deriver]
-ENABLED = false
-WORKERS = 1
-POLLING_SLEEP_INTERVAL_SECONDS = 1.0
-STALE_SESSION_TIMEOUT_MINUTES = 5
-DEDUPLICATE = true
-LOG_OBSERVATIONS = false
-
-[deriver.model_config]
-transport = "openai"
-model = "qwen3.6:27b-q4_K_M"
-
-[deriver.model_config.overrides]
-base_url = "http://ollama:11434/v1"
-api_key_env = "LLM_OPENAI_API_KEY"
-
-[summary]
-ENABLED = false
-
-[summary.model_config]
-transport = "openai"
-model = "qwen3.6:27b-q4_K_M"
-
-[summary.model_config.overrides]
-base_url = "http://ollama:11434/v1"
-api_key_env = "LLM_OPENAI_API_KEY"
-
-[dream]
-ENABLED = false
-
-[dialectic]
-MAX_OUTPUT_TOKENS = 4096
-MAX_INPUT_TOKENS = 16384
-
-[cache]
-ENABLED = false
-
-[vector_store]
-TYPE = "pgvector"
-
-[metrics]
-ENABLED = false
-
-[telemetry]
-ENABLED = false
-
-[sentry]
-ENABLED = false

env/.env.example.honcho

@@ -1,31 +0,0 @@
-# Honcho Environment Variables
-# Copy this file to your .env (at the compose root or docker-compose working directory)
-# and fill in the secrets.
-#
-#   cp env/.env.example.honcho .env
-#
-# Then reference it from compose.yml:
-#   env_file:
-#     - path: .env
-#       required: true
-
-# ---------------------------------------------------------------------------
-# Database
-# ---------------------------------------------------------------------------
-# PostgreSQL connection string for Honcho.
-# The password must match HONCHO_DB_PASSWORD below.
-HONCHO_DB_PASSWORD=change_me_to_a_strong_random_password
-
-# ---------------------------------------------------------------------------
-# LLM Provider
-# ---------------------------------------------------------------------------
-# Ollama does not require a real API key, but the env var must be set to a
-# non-empty string for the OpenAI-compatible client to connect.
-LLM_OPENAI_API_KEY=ollama
-
-# ---------------------------------------------------------------------------
-# Honcho Server
-# ---------------------------------------------------------------------------
-# Honcho will pick up DB_CONNECTION_URI from the compose environment.
-# You can override additional settings here if needed.
-# LOG_LEVEL=INFO