ci: add Gitea workflow to build fork CUDA image

Add a Gitea Actions workflow and multi-stage Containerfile that build
this fork's llama-swap (serial scheduler + embedded Svelte UI) from
source and layer it on a pinned llama.cpp CUDA server base, then push to
the Gitea container registry as v230-cuda-b9821.

- docker/fork-cuda.Containerfile: node UI -> go build -> cuda runtime,
  runs as root to match the upstream non-suffixed image
- .gitea/workflows/build-cuda-image.yml: workflow_dispatch (version +
  llama.cpp build inputs) and push-on-build-files; logs in with
  REGISTRY_USER/REGISTRY_PASSWORD

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitea/workflows/build-cuda-image.yml

@@ -0,0 +1,76 @@
+name: Build CUDA image (fork)
+
+# Builds this fork's llama-swap (serial scheduler + embedded UI) from source and
+# layers it on a pinned llama.cpp CUDA server base, then pushes to the Gitea
+# container registry, e.g. gitea.stevedudenhoeffer.com/steve/llama-swap:v230-cuda-b9821
+#
+# Requires repo secrets: REGISTRY_USER, REGISTRY_PASSWORD (push to the registry).
+
+on:
+  workflow_dispatch:
+    inputs:
+      llama_swap_version:
+        description: "llama-swap version label (image tag prefix)"
+        required: false
+        default: "v230"
+      llamacpp_build:
+        description: "llama.cpp CUDA server build (base image tag suffix)"
+        required: false
+        default: "b9821"
+  # Building the build definition itself kicks off a fresh image.
+  push:
+    branches: [main]
+    paths:
+      - ".gitea/workflows/build-cuda-image.yml"
+      - "docker/fork-cuda.Containerfile"
+
+env:
+  REGISTRY: gitea.stevedudenhoeffer.com
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Compute image metadata
+        id: meta
+        run: |
+          LS_VER="${{ inputs.llama_swap_version || 'v230' }}"
+          LCPP="${{ inputs.llamacpp_build || 'b9821' }}"
+          {
+            echo "image=${REGISTRY}/${{ github.repository }}"
+            echo "tag=${LS_VER}-cuda-${LCPP}"
+            echo "base_tag=server-cuda-${LCPP}"
+            echo "ls_version=${LS_VER}"
+            echo "build_date=$(date -u +%Y-%m-%dT%H:%M:%SZ)"
+          } >> "$GITHUB_OUTPUT"
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to Gitea registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ secrets.REGISTRY_USER }}
+          password: ${{ secrets.REGISTRY_PASSWORD }}
+
+      - name: Build and push
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: docker/fork-cuda.Containerfile
+          push: true
+          provenance: false
+          build-args: |
+            BASE_TAG=${{ steps.meta.outputs.base_tag }}
+            LS_VERSION=${{ steps.meta.outputs.ls_version }}
+            GIT_HASH=${{ github.sha }}
+            BUILD_DATE=${{ steps.meta.outputs.build_date }}
+          tags: ${{ steps.meta.outputs.image }}:${{ steps.meta.outputs.tag }}
+
+      - name: Summary
+        run: |
+          echo "Pushed ${{ steps.meta.outputs.image }}:${{ steps.meta.outputs.tag }}" >> "$GITHUB_STEP_SUMMARY"

docker/fork-cuda.Containerfile

@@ -0,0 +1,72 @@
+# Build a CUDA llama-swap image FROM THIS FORK's source (includes the serial
+# scheduler) and layer it on a pinned llama.cpp CUDA server base. Produces e.g.:
+#   gitea.stevedudenhoeffer.com/steve/llama-swap:v230-cuda-b9821
+#
+# BASE_TAG selects the llama.cpp CUDA runtime + llama-server build, e.g.
+# "server-cuda-b9821". The llama-swap binary (with the embedded Svelte UI) is
+# compiled from the repo at build time, so no GitHub release is required.
+#
+# Build context is the repo root:
+#   docker build -f docker/fork-cuda.Containerfile \
+#     --build-arg BASE_TAG=server-cuda-b9821 -t llama-swap:v230-cuda-b9821 .
+
+ARG BASE_IMAGE=ghcr.io/ggml-org/llama.cpp
+ARG BASE_TAG=server-cuda-b9821
+
+# ---- Stage 1: build the Svelte UI (embedded into the binary) ----
+FROM node:22-bookworm-slim AS ui
+WORKDIR /src/ui-svelte
+# Install deps first for layer caching.
+COPY ui-svelte/package.json ui-svelte/package-lock.json ./
+RUN npm ci
+COPY ui-svelte/ ./
+# `npm run build` is `vite build --emptyOutDir`; vite.config.ts writes to
+# ../internal/server/ui_dist, which //go:embed picks up in the next stage.
+RUN mkdir -p /src/internal/server && npm run build
+
+# ---- Stage 2: build the llama-swap binary with the embedded UI ----
+FROM golang:1.26-bookworm AS build
+WORKDIR /src
+# Cache modules independently of source churn.
+COPY go.mod go.sum ./
+RUN go mod download
+COPY . .
+# Overlay the freshly built UI so //go:embed ui_dist ships the real assets
+# instead of the committed placeholder.
+COPY --from=ui /src/internal/server/ui_dist/ ./internal/server/ui_dist/
+ARG LS_VERSION=v230
+ARG GIT_HASH=unknown
+ARG BUILD_DATE=unknown
+RUN CGO_ENABLED=0 GOOS=linux go build \
+    -ldflags="-X main.version=${LS_VERSION} -X main.commit=${GIT_HASH} -X main.date=${BUILD_DATE}" \
+    -o /out/llama-swap .
+
+# ---- Stage 3: runtime image on the pinned llama.cpp CUDA base ----
+FROM ${BASE_IMAGE}:${BASE_TAG}
+
+# Run as root by default to match the upstream `vNNN-cuda-bNNNN` (non-suffixed)
+# image that ragnaros pulls today: it needs root to reach the mounted docker
+# socket for container-backed models (sd-server). Override UID/GID at build time
+# for a non-root variant.
+ARG UID=0
+ARG GID=0
+ARG USER_HOME=/root
+ENV HOME=$USER_HOME
+
+RUN set -eux; \
+    if [ "$UID" -ne 0 ]; then \
+      if [ "$GID" -ne 0 ]; then groupadd --system --gid "$GID" app; fi; \
+      useradd --system --uid "$UID" --gid "$GID" --home "$USER_HOME" app; \
+    fi; \
+    mkdir --parents "$HOME" /app; \
+    chown --recursive "$UID:$GID" "$HOME" /app
+
+COPY --from=build --chown=$UID:$GID /out/llama-swap /app/llama-swap
+COPY --chown=$UID:$GID docker/config.example.yaml /app/config.yaml
+
+USER $UID:$GID
+WORKDIR /app
+ENV PATH="/app:${PATH}"
+
+HEALTHCHECK CMD curl -f http://localhost:8080/ || exit 1
+ENTRYPOINT [ "/app/llama-swap", "-config", "/app/config.yaml" ]