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

@@ -28,22 +28,17 @@ services:
   hermes:
     build:
       context: ./hermes
-      args:
+      ssh:
-        HERMES_PLUGIN_URLS: "git+https://code.lazyworkhorse.net/gortium/hermes-piper-plugin.git;git+https://code.lazyworkhorse.net/gortium/hermes-identity-plugin.git"
+        - default
     container_name: hermes
     entrypoint: ["/bin/bash", "-c",
-      "bash /opt/data/hermes-tools/install.sh && bash /usr/local/bin/run-multi-gateways.sh && 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
     command: gateway run
     environment:
       - OLLAMA_HOST=http://ollama:11434
-      - HERMES_DASHBOARD=1
-      # Multi-profile: comma-separated list of profiles to run as gateways.
-      # The entrypoint reads this and starts one gateway per profile.
-      # Add profiles here when they exist on disk (e.g. default,researcher,writer)
-      - HERMES_PROFILES=ashley,claire,finn,matt,paul
       - API_SERVER_ENABLED=true
       - API_SERVER_PORT=8642
       - API_SERVER_HOST=0.0.0.0
@@ -71,30 +66,6 @@ services:
       - "26"
     networks:
       - ai_backend
-      - ai_net
-    labels:
-      - "traefik.enable=true"
-      - "traefik.docker.network=ai_net"
-
-      # Router for HTTP + redirection to HTTPS
-      - "traefik.http.routers.hermes-web-http.rule=Host(`hermes.lazyworkhorse.net`)"
-      - "traefik.http.routers.hermes-web-http.entrypoints=web"
-      - "traefik.http.routers.hermes-web-http.middlewares=redirect-to-https"
-
-      # Router for HTTPS with TLS — protected by Authelia
-      - "traefik.http.routers.hermes-web-https.rule=Host(`hermes.lazyworkhorse.net`)"
-      - "traefik.http.routers.hermes-web-https.entrypoints=websecure"
-      - "traefik.http.routers.hermes-web-https.tls=true"
-      - "traefik.http.routers.hermes-web-https.tls.certresolver=njalla"
-      - "traefik.http.routers.hermes-web-https.middlewares=hermes-auth"
-
-      # Authelia forwardAuth
-      - "traefik.http.middlewares.hermes-auth.forwardauth.address=http://authelia:9091/api/verify?rd=https://auth.lazyworkhorse.net/"
-      - "traefik.http.middlewares.hermes-auth.forwardauth.trustforwardheader=true"
-      - "traefik.http.middlewares.hermes-auth.forwardauth.authresponseheaders=X-Forwarded-User,X-Forwarded-Groups"
-
-      # Service Loadbalancer (dashboard port 9119)
-      - "traefik.http.services.hermes-web.loadbalancer.server.port=9119"
 
   syncthing:
     image: syncthing/syncthing:latest
@@ -141,22 +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:
@@ -166,47 +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/Dockerfile

@@ -1,22 +1,45 @@
 # syntax=docker/dockerfile:1
-# Hermes Agent -- official image + custom plugins layered on top.
+# Hermes Agent -- custom fork build
-# No fork needed — customizations are pip-installable plugins from Gitea.
+# Builds on top of official image + overlays our forked source from Gitea.
+# Requires Docker BuildKit. Pass SSH agent for git clone:
 #   docker compose build hermes
 # Or manually:
-#   DOCKER_BUILDKIT=1 docker build --build-arg HERMES_PLUGIN_URLS="url1 url2" -t hermes-agent:custom .
+#   DOCKER_BUILDKIT=1 docker build --ssh default -t hermes-agent:custom .
 
 # ---------- Base: official Hermes image (system deps, npm, uv, Playwright) ----------
 FROM nousresearch/hermes-agent:latest
 
-# ---------- Plugin URLs (semicolon-separated, set via compose.yml build args) ----------
+# ---------- Overlay our forked source ----------
-ARG HERMES_PLUGIN_URLS=""
+# Uses SSH agent forwarding from the build host (no key baked into image).
+# --exclude node_modules/.venv keeps the base image's pre-built layers intact.
+# Only the Python source, web UI source, and config change.
+RUN --mount=type=ssh \
+    mkdir -p /root/.ssh && \
+    ssh-keyscan -p 2222 code.lazyworkhorse.net >> /root/.ssh/known_hosts 2>/dev/null && \
+    cd /tmp && \
+    GIT_SSH_COMMAND='ssh -p 2222 -o StrictHostKeyChecking=no' \
+    git clone --depth 1 --branch main \
+    git@code.lazyworkhorse.net:gortium/hermes-agent.git fork && \
+    rsync -a --delete fork/ /opt/hermes/ \
+      --exclude node_modules \
+      --exclude .venv \
+      --exclude .git && \
+    rm -rf /tmp/fork /root/.ssh/
+
+# ---------- Rebuild web UI ----------
+# Source files changed; node_modules (from base image) reused.
+RUN cd /opt/hermes && npm run build
+
+# ---------- Reinstall Python package (editable) ----------
+# Picks up source changes from our fork.
+RUN . /opt/hermes/.venv/bin/activate && \
+    uv pip install --no-cache-dir --no-deps -e /opt/hermes
 
 # ---------- Extra system deps ----------
 USER root
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
         libportaudio2 ca-certificates poppler-utils imagemagick \
-        libolm-dev \
         texlive-latex-base texlive-latex-extra texlive-fonts-recommended \
         texlive-xetex texlive-science \
         qemu-user-static binfmt-support emacs-nox && \
@@ -25,10 +48,6 @@ RUN apt-get update && \
 # ---------- UV ----------
 COPY --chmod=0755 --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/
 
-# ---------- Matrix bridge + extra pip deps ----------
-RUN . /opt/hermes/.venv/bin/activate && \
-    uv pip install --no-cache-dir 'mautrix[encryption]' openai
-
 # ---------- Piper TTS ----------
 RUN . /opt/hermes/.venv/bin/activate && \
     uv pip install --no-cache-dir piper-tts sounddevice numpy && \
@@ -56,22 +75,9 @@ os.remove(tgz)
 print('himalaya v1.2.0 installed')
 PYEOF
 
-# ---------- Install custom plugins from URLs ----------
+# ---------- Install himalaya-ro wrapper ----------
-# HERMES_PLUGIN_URLS is a semicolon-separated list of pip-installable
+COPY --chmod=0755 himalaya-ro.sh /usr/local/bin/himalaya-ro
-# package URLs (e.g. git+https:// or direct .tar.gz archives from Gitea).
-# Each plugin is installed into the Hermes venv.
-RUN if [ -n "$HERMES_PLUGIN_URLS" ]; then \
-        . /opt/hermes/.venv/bin/activate && \
-        IFS=';' read -ra URLS <<< "$HERMES_PLUGIN_URLS" && \
-        for url in "${URLS[@]}"; do \
-            echo "Installing plugin: $url" && \
-            uv pip install --no-cache-dir "$url"; \
-        done; \
-    fi
 
-# ---------- Install multi-gateway launcher ----------
-# Launches one gateway process per profile (HERMES_PROFILES env var)
-COPY --chmod=0755 run-multi-gateways.sh /usr/local/bin/run-multi-gateways.sh
 
 # ---------- Runtime ----------
 USER hermes
@@ -82,7 +88,6 @@ ENV CHROME_EXECUTABLE=/opt/hermes/.playwright/chromium/chrome-linux/chrome
 
 # Ensure tools directory and toolsets.py are writable by the hermes runtime user
 # so custom tools can be injected from the persistent volume at startup.
-USER root
 RUN chown -R hermes:hermes /opt/hermes/tools /opt/hermes/toolsets.py
 
-VOLUME [ "/opt/data" ]
+VOLUME [ "/opt/data" ]

ai/hermes/himalaya-ro.sh

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+# ─────────────────────────────────────────────────────────────
+# himalaya-ro — Read-only wrapper for himalaya
+#
+# Blocks destructive commands and logs audit trail.
+# Pass-through for read-only commands (list, read, search).
+#
+# Usage:  himalaya-ro [options] <command> [args...]
+#
+# Install: place in PATH before the real himalaya, or use
+#          `ln -sf himalaya-ro /usr/local/bin/himalaya`
+# ─────────────────────────────────────────────────────────────
+set -o pipefail
+
+# ── Configuration ───────────────────────────────────────────
+HIMALAYA_BIN="${HIMALAYA_BIN:-/usr/local/bin/himalaya}"
+AUDIT_LOG="${HIMALAYA_AUDIT_LOG:-/var/log/himalaya-audit.log}"
+
+# ── Destructive commands we block ──────────────────────────
+BLOCKED_CMDS=(
+  "message move"
+  "message delete"
+  "message copy"
+  "flag add"
+  "flag remove"
+  "folder create"
+  "folder delete"
+  "folder rename"
+  "template send"
+  "account configure"
+  "account delete"
+)
+
+# ── Determine the subcommand being invoked ─────────────────
+# Strip leading options (--account, --output, etc.) to find the verb
+ARGS=()
+SKIP_NEXT=false
+for arg in "$@"; do
+  if $SKIP_NEXT; then
+    SKIP_NEXT=false
+    continue
+  fi
+  if [[ "$arg" == --* ]]; then
+    case "$arg" in
+      --account|--output|--page|--page-size|--folder|--color|--format)
+        SKIP_NEXT=true ;;
+    esac
+    continue
+  fi
+  ARGS+=("$arg")
+done
+
+# Build subcommand string and check against blocklist
+CMD_STR=""
+for ((i=0; i<${#ARGS[@]}; i++)); do
+  if [ -z "$CMD_STR" ]; then
+    CMD_STR="${ARGS[$i]}"
+  else
+    CMD_STR="$CMD_STR ${ARGS[$i]}"
+  fi
+  for blocked in "${BLOCKED_CMDS[@]}"; do
+    if [[ "$CMD_STR" == "$blocked" ]]; then
+      TS=$(date '+%Y-%m-%d %H:%M:%S')
+      echo "[AUDIT] $TS BLOCKED: himalaya $*" >> "$AUDIT_LOG"
+      echo "ERROR: Command 'himalaya $CMD_STR ...' is blocked by read-only policy." >&2
+      echo "       Audit log: $AUDIT_LOG" >&2
+      exit 100
+    fi
+  done
+done
+
+# ── Allow pass-through ─────────────────────────────────────
+exec "$HIMALAYA_BIN" "$@"

ai/hermes/run-multi-gateways.sh

@@ -1,32 +0,0 @@
-#!/bin/bash
-# Multi-gateway launcher for HERMES_PROFILES env var.
-# Reads comma-separated profile names, spawns one gateway per profile.
-# Designed to run before the main entrypoint — gateways run in background.
-set -e
-
-if [ -z "${HERMES_PROFILES}" ]; then
-  echo "HERMES_PROFILES not set — skipping multi-gateway launch"
-  exit 0
-fi
-
-# Source venv to make 'hermes' available (entrypoint.sh sources it later,
-# but we need it NOW for the background gateways)
-HERMES_BIN="/opt/hermes/.venv/bin/hermes"
-if [ ! -x "$HERMES_BIN" ]; then
-  echo "ERROR: hermes binary not found at $HERMES_BIN"
-  exit 1
-fi
-
-mkdir -p /opt/data/logs
-
-IFS=',' read -ra PROFILES <<< "${HERMES_PROFILES}"
-for profile in "${PROFILES[@]}"; do
-  profile="$(echo "${profile}" | xargs)"  # trim whitespace
-  [ -z "${profile}" ] && continue
-
-  echo "Starting gateway for profile: ${profile}"
-  nohup env API_SERVER_ENABLED=false API_SERVER_KEY= gosu hermes "$HERMES_BIN" --profile "${profile}" gateway run \
-      >> "/opt/data/logs/gateway-${profile}.log" 2>&1 &
-done
-
-echo "All gateways launched: ${HERMES_PROFILES}"

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"]