Spaces:

Benaah
/

MicProtocol

Running

Upload folder using huggingface_hub

6abfd9d verified about 21 hours ago

6.9 kB

# ═══════════════════════════════════════════════════════════
# GuestFlow API - Production Dockerfile
# Multi-stage build for optimized production image
# ═══════════════════════════════════════════════════════════

# ─── Stage 1: Pruner ───────────────────────────────────────
FROM node:20-slim AS pruner
RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev sqlite3 libsqlite3-dev g++ make python3 && rm -rf /var/lib/apt/lists/*
WORKDIR /app
RUN npm install -g turbo@2.8.12
COPY . .
RUN turbo prune --scope=@micprotocol/api --docker

# ─── Stage 2: Dependencies ─────────────────────────────────
FROM node:20-slim AS deps
RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev sqlite3 libsqlite3-dev g++ make python3 openssl && rm -rf /var/lib/apt/lists/*
WORKDIR /app

# Copy pruned package.json and lockfile
COPY --from=pruner /app/out/json/ .

# Install pnpm
RUN npm install -g pnpm@10.30.3

# Install dependencies
RUN pnpm install --frozen-lockfile

# ─── Stage 3: Builder ──────────────────────────────────────
FROM node:20-slim AS builder
RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev sqlite3 libsqlite3-dev g++ make python3 openssl && rm -rf /var/lib/apt/lists/*
WORKDIR /app

# Install pnpm
RUN npm install -g pnpm@10.30.3

# Copy dependencies from deps stage (includes all node_modules, lockfile, and package.json files)
COPY --from=deps /app /app

# Copy pruned source files on top of it
COPY --from=pruner /app/out/full/ .

# Generate Prisma client
RUN cd apps/api && npx prisma generate

# Build application
RUN NODE_OPTIONS="--max-old-space-size=4096" pnpm build --filter @micprotocol/api

# Prune devDependencies to keep only production dependencies
RUN CI=true pnpm install --prod --frozen-lockfile

# ─── Stage 4: Production ─────────────────────────────────
FROM node:20-slim AS runner
WORKDIR /app

ENV NODE_ENV=production
ENV PORT=4000
ENV MEILISEARCH_URL=http://localhost:7700
ENV MEILISEARCH_API_KEY=masterKey

RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev openssl libgcc1 tini ffmpeg sqlite3 libsqlite3-dev netcat-openbsd python3 && rm -rf /var/lib/apt/lists/*

# Copy Meilisearch binary from the official image
COPY --from=getmeili/meilisearch:v1.9.0 /bin/meilisearch /usr/local/bin/meilisearch

# Create non-root user
RUN addgroup --system --gid 1001 nodejs
RUN adduser --system --uid 1001 nestjs

# Copy entire pruned application and dependencies
COPY --from=builder --chown=nestjs:nodejs /app /app

# ─── Python clip-scoring engine ──────────────────────────
# The BullMQ clip-scoring processor spawns
# `python3 services/clip-scorer/pipeline.py` via execFile
# (apps/api/src/queue/clip-scoring.processor.ts), so this image must
# ship python3, the ML stack, and the pipeline source. torch is pinned
# to the CPU wheel index: CUDA is unnecessary for the 224px ViT-B/32
# inference and would add ~2GB of NVIDIA runtime deps.
# The venv is pinned to the system python: without it uv probes the
# newest CPython it can download (3.15 here), for which torchvision has
# no wheels — the clip-scoring stack then can't resolve. torch AND
# torchvision must land in one resolution pass from the CPU index:
# open-clip-torch declares both as deps, and a PyPI-resolved torchvision
# breaks open_clip import with "operator torchvision::nms does not
# exist" (same failure class as the standalone clip-scorer image).
# Build with the monorepo root as context (docker-compose does exactly
# that for the `api` and `worker` services) so services/clip-scorer is
# in scope — turbo prune does not carry it (not a pnpm workspace).
COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
COPY --chown=nestjs:nodejs services/clip-scorer/ /app/services/clip-scorer/
ENV PATH="/opt/venv/bin:$PATH" \
    HF_HOME=/app/.hf-cache
RUN uv venv --python /usr/bin/python3 /opt/venv \
    && VIRTUAL_ENV=/opt/venv uv pip install --index-url https://download.pytorch.org/whl/cpu "torch>=2.0.0" "torchvision>=0.15.0" \
    && VIRTUAL_ENV=/opt/venv uv pip install "open-clip-torch>=2.20.0" "transformers>=4.30.0" "Pillow>=10.0.0" "soundfile>=0.12.0" "pyloudnorm>=0.1.0" "numpy>=1.24.0"
# Pre-baked CLIP weights come from the standalone clip-scorer image
# (its build already downloaded the model) so this build does NOT need
# Hub access. Placed after the venv layer so upgrading torch does not
# force a weight re-copy (and vice versa).
COPY --from=localhost/micprotocol-clip-scorer:audit-verify \
     /root/.cache/huggingface/hub/ \
     /app/.hf-cache/hub/
# Offline verification: fail the build if the baked cache ever stops
# matching what open_clip 3.x expects to find. BART (transcript NLI) is
# intentionally not baked — it is a ~1.6GB download and the pipeline
# falls back to the heuristic scorer when it cannot download at runtime.
RUN HF_HUB_OFFLINE=1 python3 -c "import open_clip; open_clip.create_model_and_transforms('ViT-B-32', pretrained='laion2b_s34b_b79k')"

# Create data directory for Meilisearch and set ownership
RUN mkdir -p /app/meili_data && chown -R nestjs:nodejs /app

# Switch to non-root user
USER nestjs

# Create startup script
RUN echo '#!/bin/sh' > /app/start.sh && \
    echo 'MEILI_HOST=$(echo ${MEILISEARCH_HOST:-localhost} | sed "s|http://||;s|https://||")' >> /app/start.sh && \
    echo 'MEILI_PORT=$(echo $MEILI_HOST | cut -d: -f2)' >> /app/start.sh && \
    echo '[ -z "$MEILI_PORT" ] && MEILI_PORT=7700' >> /app/start.sh && \
    echo 'MEILI_HOST=$(echo $MEILI_HOST | cut -d: -f1)' >> /app/start.sh && \
    echo 'echo "Waiting for Meilisearch at $MEILI_HOST:$MEILI_PORT..."' >> /app/start.sh && \
    echo 'while ! nc -z $MEILI_HOST $MEILI_PORT; do' >> /app/start.sh && \
    echo '  sleep 0.5' >> /app/start.sh && \
    echo 'done' >> /app/start.sh && \
    echo 'echo "Meilisearch is ready!"' >> /app/start.sh && \
    echo 'node /app/apps/api/dist/main.js' >> /app/start.sh && \
    chmod +x /app/start.sh

# Expose ports
EXPOSE 4000
EXPOSE 7700

# Health check
HEALTHCHECK --interval=30s --timeout=3s --start-period=10s --retries=3 \
    CMD node -e "require('http').get('http://localhost:4000/health', (r) => r.statusCode === 200 ? process.exit(0) : process.exit(1))"

# Start application
ENTRYPOINT ["tini", "--"]
CMD ["/app/start.sh"]

	# ═══════════════════════════════════════════════════════════
	# GuestFlow API - Production Dockerfile
	# Multi-stage build for optimized production image
	# ═══════════════════════════════════════════════════════════

	# ─── Stage 1: Pruner ───────────────────────────────────────
	FROM node:20-slim AS pruner
	RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev sqlite3 libsqlite3-dev g++ make python3 && rm -rf /var/lib/apt/lists/*
	WORKDIR /app
	RUN npm install -g turbo@2.8.12
	COPY . .
	RUN turbo prune --scope=@micprotocol/api --docker

	# ─── Stage 2: Dependencies ─────────────────────────────────
	FROM node:20-slim AS deps
	RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev sqlite3 libsqlite3-dev g++ make python3 openssl && rm -rf /var/lib/apt/lists/*
	WORKDIR /app

	# Copy pruned package.json and lockfile
	COPY --from=pruner /app/out/json/ .

	# Install pnpm
	RUN npm install -g pnpm@10.30.3

	# Install dependencies
	RUN pnpm install --frozen-lockfile

	# ─── Stage 3: Builder ──────────────────────────────────────
	FROM node:20-slim AS builder
	RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev sqlite3 libsqlite3-dev g++ make python3 openssl && rm -rf /var/lib/apt/lists/*
	WORKDIR /app

	# Install pnpm
	RUN npm install -g pnpm@10.30.3

	# Copy dependencies from deps stage (includes all node_modules, lockfile, and package.json files)
	COPY --from=deps /app /app

	# Copy pruned source files on top of it
	COPY --from=pruner /app/out/full/ .

	# Generate Prisma client
	RUN cd apps/api && npx prisma generate

	# Build application
	RUN NODE_OPTIONS="--max-old-space-size=4096" pnpm build --filter @micprotocol/api

	# Prune devDependencies to keep only production dependencies
	RUN CI=true pnpm install --prod --frozen-lockfile

	# ─── Stage 4: Production ─────────────────────────────────
	FROM node:20-slim AS runner
	WORKDIR /app

	ENV NODE_ENV=production
	ENV PORT=4000
	ENV MEILISEARCH_URL=http://localhost:7700
	ENV MEILISEARCH_API_KEY=masterKey

	RUN apt-get update && apt-get install -y --no-install-recommends libc6-dev openssl libgcc1 tini ffmpeg sqlite3 libsqlite3-dev netcat-openbsd python3 && rm -rf /var/lib/apt/lists/*

	# Copy Meilisearch binary from the official image
	COPY --from=getmeili/meilisearch:v1.9.0 /bin/meilisearch /usr/local/bin/meilisearch

	# Create non-root user
	RUN addgroup --system --gid 1001 nodejs
	RUN adduser --system --uid 1001 nestjs

	# Copy entire pruned application and dependencies
	COPY --from=builder --chown=nestjs:nodejs /app /app

	# ─── Python clip-scoring engine ──────────────────────────
	# The BullMQ clip-scoring processor spawns
	# `python3 services/clip-scorer/pipeline.py` via execFile
	# (apps/api/src/queue/clip-scoring.processor.ts), so this image must
	# ship python3, the ML stack, and the pipeline source. torch is pinned
	# to the CPU wheel index: CUDA is unnecessary for the 224px ViT-B/32
	# inference and would add ~2GB of NVIDIA runtime deps.
	# The venv is pinned to the system python: without it uv probes the
	# newest CPython it can download (3.15 here), for which torchvision has
	# no wheels — the clip-scoring stack then can't resolve. torch AND
	# torchvision must land in one resolution pass from the CPU index:
	# open-clip-torch declares both as deps, and a PyPI-resolved torchvision
	# breaks open_clip import with "operator torchvision::nms does not
	# exist" (same failure class as the standalone clip-scorer image).
	# Build with the monorepo root as context (docker-compose does exactly
	# that for the `api` and `worker` services) so services/clip-scorer is
	# in scope — turbo prune does not carry it (not a pnpm workspace).
	COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
	COPY --chown=nestjs:nodejs services/clip-scorer/ /app/services/clip-scorer/
	ENV PATH="/opt/venv/bin:$PATH" \
	HF_HOME=/app/.hf-cache
	RUN uv venv --python /usr/bin/python3 /opt/venv \
	&& VIRTUAL_ENV=/opt/venv uv pip install --index-url https://download.pytorch.org/whl/cpu "torch>=2.0.0" "torchvision>=0.15.0" \
	&& VIRTUAL_ENV=/opt/venv uv pip install "open-clip-torch>=2.20.0" "transformers>=4.30.0" "Pillow>=10.0.0" "soundfile>=0.12.0" "pyloudnorm>=0.1.0" "numpy>=1.24.0"
	# Pre-baked CLIP weights come from the standalone clip-scorer image
	# (its build already downloaded the model) so this build does NOT need
	# Hub access. Placed after the venv layer so upgrading torch does not
	# force a weight re-copy (and vice versa).
	COPY --from=localhost/micprotocol-clip-scorer:audit-verify \
	/root/.cache/huggingface/hub/ \
	/app/.hf-cache/hub/
	# Offline verification: fail the build if the baked cache ever stops
	# matching what open_clip 3.x expects to find. BART (transcript NLI) is
	# intentionally not baked — it is a ~1.6GB download and the pipeline
	# falls back to the heuristic scorer when it cannot download at runtime.
	RUN HF_HUB_OFFLINE=1 python3 -c "import open_clip; open_clip.create_model_and_transforms('ViT-B-32', pretrained='laion2b_s34b_b79k')"

	# Create data directory for Meilisearch and set ownership
	RUN mkdir -p /app/meili_data && chown -R nestjs:nodejs /app

	# Switch to non-root user
	USER nestjs

	# Create startup script
	RUN echo '#!/bin/sh' > /app/start.sh && \
	echo 'MEILI_HOST=$(echo ${MEILISEARCH_HOST:-localhost} \| sed "s\|http://\|\|;s\|https://\|\|")' >> /app/start.sh && \
	echo 'MEILI_PORT=$(echo $MEILI_HOST \| cut -d: -f2)' >> /app/start.sh && \
	echo '[ -z "$MEILI_PORT" ] && MEILI_PORT=7700' >> /app/start.sh && \
	echo 'MEILI_HOST=$(echo $MEILI_HOST \| cut -d: -f1)' >> /app/start.sh && \
	echo 'echo "Waiting for Meilisearch at $MEILI_HOST:$MEILI_PORT..."' >> /app/start.sh && \
	echo 'while ! nc -z $MEILI_HOST $MEILI_PORT; do' >> /app/start.sh && \
	echo ' sleep 0.5' >> /app/start.sh && \
	echo 'done' >> /app/start.sh && \
	echo 'echo "Meilisearch is ready!"' >> /app/start.sh && \
	echo 'node /app/apps/api/dist/main.js' >> /app/start.sh && \
	chmod +x /app/start.sh

	# Expose ports
	EXPOSE 4000
	EXPOSE 7700

	# Health check
	HEALTHCHECK --interval=30s --timeout=3s --start-period=10s --retries=3 \
	CMD node -e "require('http').get('http://localhost:4000/health', (r) => r.statusCode === 200 ? process.exit(0) : process.exit(1))"

	# Start application
	ENTRYPOINT ["tini", "--"]
	CMD ["/app/start.sh"]