implementing ROCM and metal support

.env.discord.example

@@ -6,6 +6,7 @@ Discord_allowed_users=
 # Explicit server TEXT channel IDs; an empty list disables server interaction.
 # DMs are allowed only for configured admins/trusted users.
 Discord_Channel_whitelist=234567890123456789
+# Only for a client started by hand: Start Discord (Settings or the tray) passes the backend's own address.
 Discord_backend_url=http://127.0.0.1:8765
 # Optional test server for immediate slash-command synchronization.
 Discord_sync_guild=

.gitattributes

@@ -1,2 +1,4 @@
 # Auto detect text files and perform LF normalization
 * text=auto
+# Shell scripts must keep LF on Windows checkouts, or bash rejects "set -o pipefail\r".
+*.sh text eol=lf

.github/actions/hip-bundle/action.yml

@@ -0,0 +1,42 @@
+name: HIP bundle
+description: >-
+  The opt-in hip bundle (electron/native_backends.json), built by tools/release/build.py --smoke-bundle exactly as a release
+  builds a bundle, inside the job's rocm/dev-ubuntu-22.04 container, then checked there without a GPU: the rpath that finds
+  AMD's ROCm, and build.py --check-bundle (the six exports and riko_devices' device list, a CPU module from the bundle,
+  ggml-hip loading against ROCm 7.2.1). rocm-compile.yml's compile job and release.yml's opt-in hip job both run it, so
+  the compile job proves the release's recipe; neither runs a model on an AMD GPU.
+inputs:
+  destination:
+    description: The bundle folder, relative to the workspace and named hip (container jobs see runner.temp under another path)
+    required: true
+runs:
+  using: composite
+  steps:
+    - name: ROCm build dependencies
+      # The image has ROCm 7.2.1's compilers and Python 3.10, and its apt source is that release's repository, signed by AMD.
+      shell: bash
+      run: |
+        apt-get update
+        DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends git cmake hipblas-dev rocblas-dev
+        git config --global --add safe.directory '*'
+    - name: llama.cpp at the pin
+      shell: bash
+      run: |
+        pin=$(python3 -c "import runpy; print(runpy.run_path('tools/release/build.py')['PIN'])")
+        git init -q "$RUNNER_TEMP/llama.cpp-hip"
+        git -C "$RUNNER_TEMP/llama.cpp-hip" remote add origin https://github.com/ggml-org/llama.cpp
+        git -C "$RUNNER_TEMP/llama.cpp-hip" fetch -q --depth 1 origin "$pin"
+        git -C "$RUNNER_TEMP/llama.cpp-hip" -c core.autocrlf=false checkout -q --detach FETCH_HEAD
+    - name: Build the hip bundle
+      # CMake's HIP language takes ROCm's clang from HIPCXX, as llama.cpp's docs build it; ggml-hip for twelve GPU targets
+      # is the long part, so every core compiles.
+      shell: bash
+      run: |
+        export HIPCXX="$(/opt/rocm/bin/hipconfig -l)/clang" HIP_PATH="$(/opt/rocm/bin/hipconfig -R)" RIKO_BUILD_JOBS="$(nproc)"
+        python3 tools/release/build.py --smoke-bundle hip "$RUNNER_TEMP/llama.cpp-hip" "${{ inputs.destination }}"
+    - name: Check the hip bundle without a GPU
+      # The manifest's rpath: ggml-hip looks beside itself, then in /opt/rocm/lib, where AMD's packages put libamdhip64.
+      shell: bash
+      run: |
+        readelf -d "${{ inputs.destination }}/libggml-hip.so" | grep -E 'Library (runpath|rpath): \[\$ORIGIN:/opt/rocm/lib\]'
+        python3 tools/release/build.py --check-bundle "${{ inputs.destination }}/libriko-native.so"

.github/workflows/ci.yml

@@ -0,0 +1,83 @@
+name: CI
+# Every push and pull request: the whole Python suite on each OS a release ships for, and the Electron tests and build.
+# Every GPU, audio device and native library is mocked here; .github/workflows/native-smoke.yml loads a real build.
+on:
+  push:
+    branches: ['**']  # not tags: a v* tag runs the release workflow
+  pull_request:
+  workflow_dispatch:
+permissions:
+  contents: read
+concurrency:
+  # A newer push to the same branch or pull request replaces a run still in progress.
+  group: ${{ github.workflow }}-${{ github.head_ref || github.ref_name }}  # a push and its pull request share one run
+  cancel-in-progress: true
+jobs:
+  python:
+    name: Python tests (${{ matrix.os }})
+    strategy:
+      fail-fast: false
+      matrix:
+        # The release workflow's runners: Windows and Linux x64, Apple silicon macOS.
+        os: [ubuntu-22.04, windows-2022, macos-15]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 30
+    defaults:
+      run:
+        # bash -eo pipefail stops a step at its first failing command; Windows' default pwsh checks only the last one.
+        shell: bash
+    env:
+      # No Hugging Face downloads: tests/conftest.py already refuses Julia-1's, and this catches any other model.
+      HF_HUB_OFFLINE: '1'
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065  # v5.6.0
+        with:
+          python-version: '3.11'
+          cache: pip
+          cache-dependency-path: tools/release/requirements-lock*.txt
+      - name: Linux audio libraries
+        # sounddevice loads PortAudio when imported (tests/test_release_build.py runs the release check's imports).
+        if: runner.os == 'Linux'
+        run: sudo apt-get update && sudo apt-get install -y libportaudio2 libsndfile1
+      - name: Python dependencies
+        # The release's hash-locked versions (tools/release/lock.py), so CI tests what a release freezes; torch is the CPU build.
+        run: |
+          python -m pip install --require-hashes -r tools/release/requirements-lock.txt --extra-index-url https://download.pytorch.org/whl/cpu
+          python -m pip install --require-hashes --no-deps -r tools/release/requirements-lock-no-deps.txt
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4.4.0
+        with:
+          node-version: '22'
+          cache: npm
+          cache-dependency-path: electron/package-lock.json
+      - name: Electron packages for the cross-language tests
+        # tests/test_settings_store.py runs electron/release.cjs's first-run setup in Node and validates what it writes; it
+        # skips without electron/node_modules. Only the JavaScript packages are needed, not the Electron binary.
+        working-directory: electron
+        run: npm ci --ignore-scripts
+      - name: Python tests
+        # --timeout (pytest-timeout, in the lock) fails a deadlocked test after 5 minutes instead of hanging the job;
+        # the whole suite takes well under a minute on a developer machine.
+        run: python -m pytest -q --timeout=300 -rfEs
+
+  electron:
+    name: Electron tests and build
+    runs-on: ubuntu-22.04
+    timeout-minutes: 20
+    defaults:
+      run:
+        working-directory: electron
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4.4.0
+        with:
+          node-version: '22'
+          cache: npm
+          cache-dependency-path: electron/package-lock.json
+      - run: npm ci
+      - name: Main and preload syntax
+        run: |
+          node --check main.cjs
+          node --check preload.cjs
+      - run: npm test
+      - run: npm run build

.github/workflows/native-smoke.yml

@@ -0,0 +1,120 @@
+name: Native smoke
+# Builds the patched riko-native bridge at the pinned llama.cpp with Metal, as a release does, and runs a small GGUF through
+# InProcessLlamaProvider (tests/test_native_smoke.py): the ctypes ABI, slots, streaming, cancellation and the emotion probe.
+# Hosted Apple silicon runners have a paravirtual GPU that llama.cpp's Metal backend can use (PyTorch MPS cannot).
+# The opt-in ROCm bundle's compile check lives in rocm-compile.yml (manual: twelve GPU targets take hours).
+on:
+  push:
+    branches: ['**']
+    paths:
+      - tools/llama_cpp/**
+      - Code/process/app_core/inference/**
+      - Code/process/app_core/kernel/**   # messages, streaming, cancellation and timings the providers use
+      - tools/release/**
+      - electron/native_backends.json
+      - Code/process/app_core/emotion/**   # FEATURE_VERSION must match tools/llama_cpp/emotion-probe.patch
+      - Code/process/app_core/configuration/**
+      - Code/run_server.py
+      - pyproject.toml
+      - tests/test_native_smoke.py
+      - .github/workflows/native-smoke.yml
+  pull_request:
+    paths:  # the same list as push
+      - tools/llama_cpp/**
+      - Code/process/app_core/inference/**
+      - Code/process/app_core/kernel/**   # messages, streaming, cancellation and timings the providers use
+      - tools/release/**
+      - electron/native_backends.json
+      - Code/process/app_core/emotion/**   # FEATURE_VERSION must match tools/llama_cpp/emotion-probe.patch
+      - Code/process/app_core/configuration/**
+      - Code/run_server.py
+      - pyproject.toml
+      - tests/test_native_smoke.py
+      - .github/workflows/native-smoke.yml
+  schedule:
+    - cron: '23 5 * * 1'  # weekly (Monday 05:23 UTC): runner images, Xcode and macOS change without a commit here
+  workflow_dispatch:
+permissions:
+  contents: read
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.ref_name }}  # a push and its pull request share one run
+  cancel-in-progress: true
+jobs:
+  metal:
+    name: Metal bridge and model (macos-15 arm64)
+    runs-on: macos-15
+    timeout-minutes: 60
+    defaults:
+      run:
+        shell: bash
+    env:
+      HF_HUB_OFFLINE: '1'
+      LLAMA_CPP: ${{ github.workspace }}/.native/llama.cpp-smoke
+      # Qwen2.5-0.5B-Instruct, Apache-2.0 (LICENSE in the repository at this revision), from Qwen's own GGUF repository.
+      MODEL_URL: https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/9217f5db79a29953eb74d5343926648285ec7e67/qwen2.5-0.5b-instruct-q4_k_m.gguf
+      MODEL_SHA256: 74a4da8c9fdbcd15bd1f6d01d621410d31c6fc00986f5eb687824e7b93d7a9db
+      MODEL: ${{ github.workspace }}/.native/models/qwen2.5-0.5b-instruct-q4_k_m.gguf
+    steps:
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065  # v5.6.0
+        with:
+          python-version: '3.11'
+          cache: pip
+          cache-dependency-path: tools/release/requirements-lock*.txt
+      - name: Pinned llama.cpp commit
+        id: pin
+        run: echo "commit=$(python -c "import runpy; print(runpy.run_path('tools/release/build.py')['PIN'])")" >> "$GITHUB_OUTPUT"
+      - name: Restore the llama.cpp checkout
+        id: llama-cache
+        uses: actions/cache/restore@0057852bfaa89a56745cba8c7296529d2fc39830  # v4.3.0
+        with:
+          path: .native/llama.cpp-smoke
+          key: llama.cpp-${{ steps.pin.outputs.commit }}
+      - name: Fetch llama.cpp at the pin
+        if: steps.llama-cache.outputs.cache-hit != 'true'
+        run: |
+          git init -q "$LLAMA_CPP"
+          git -C "$LLAMA_CPP" remote add origin https://github.com/ggml-org/llama.cpp
+          git -C "$LLAMA_CPP" fetch --depth 1 origin "${{ steps.pin.outputs.commit }}"
+          git -C "$LLAMA_CPP" -c core.autocrlf=false checkout -q --detach FETCH_HEAD
+      - name: Cache the unpatched checkout
+        # Saved before the patch and the build, so every run patches the pin afresh; build.py refuses any other commit.
+        if: steps.llama-cache.outputs.cache-hit != 'true'
+        uses: actions/cache/save@0057852bfaa89a56745cba8c7296529d2fc39830  # v4.3.0
+        with:
+          path: .native/llama.cpp-smoke
+          key: llama.cpp-${{ steps.pin.outputs.commit }}
+      - name: Build the Metal bundle
+        # The release's patch, CMake flags (@loader_path rpath, embedded Metal library) and library copying.
+        run: python tools/release/build.py --smoke-bundle metal "$LLAMA_CPP" "$RUNNER_TEMP/native/metal"
+      - name: Restore the model
+        id: model-cache
+        uses: actions/cache/restore@0057852bfaa89a56745cba8c7296529d2fc39830  # v4.3.0
+        with:
+          path: .native/models
+          key: gguf-${{ env.MODEL_SHA256 }}
+      - name: Download the model
+        if: steps.model-cache.outputs.cache-hit != 'true'
+        run: |
+          mkdir -p "$(dirname "$MODEL")"
+          curl --fail --location --retry 3 --output "$MODEL" "$MODEL_URL"
+      - name: Verify the model
+        run: echo "$MODEL_SHA256  $MODEL" | shasum -a 256 -c -
+      - name: Cache the model
+        # Once verified, and whether or not the tests below pass.
+        if: steps.model-cache.outputs.cache-hit != 'true'
+        uses: actions/cache/save@0057852bfaa89a56745cba8c7296529d2fc39830  # v4.3.0
+        with:
+          path: .native/models
+          key: gguf-${{ env.MODEL_SHA256 }}
+      - name: Python dependencies
+        run: |
+          python -m pip install --require-hashes -r tools/release/requirements-lock.txt --extra-index-url https://download.pytorch.org/whl/cpu
+          python -m pip install --require-hashes --no-deps -r tools/release/requirements-lock-no-deps.txt
+      - name: Native tests
+        env:
+          RIKO_TEST_NATIVE_LIBRARY: ${{ runner.temp }}/native/metal/libriko-native.dylib
+          RIKO_TEST_GGUF: ${{ env.MODEL }}
+          RIKO_TEST_EXPECT_GPU: '1'
+          RIKO_TEST_EXPECT_DEVICES: '1'   # the bridge built here has riko_devices
+        run: python -m pytest -m native -v -rs --timeout=600 tests/test_native_smoke.py