standalone build

build/Dockerfile

@@ -1,12 +1,27 @@
-FROM python:3.13-slim
+# BiRefNet background removal service — installs the published package on the CUDA 12.6 runtime.
+# Same CUDA base family as the main ../Dockerfile, but installs rmbg-as-a-service from TestPyPI
+# with a plain pip rather than building from the local source tree.
+FROM nvidia/cuda:12.6.1-cudnn-runtime-ubuntu24.04
 
 ENV DEBIAN_FRONTEND=noninteractive \
     PYTHONUNBUFFERED=1 \
+    PIP_BREAK_SYSTEM_PACKAGES=1 \
     HF_HOME=/app/hf_cache \
     HF_HUB_ENABLE_HF_TRANSFER=1 \
     PORT=8000
 
-RUN pip install rmbg-as-a-service hf-transfer
+# Ubuntu 24.04 ships Python 3.12 (the project requires >=3.12).
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends python3 python3-pip ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+# rmbg-as-a-service from TestPyPI; its deps from real PyPI; CUDA torch from the PyTorch index.
+# pip prefers torch's +cu126 build because the local version label outranks the plain wheel.
+RUN pip install \
+    --index-url https://test.pypi.org/simple/ \
+    --extra-index-url https://pypi.org/simple/ \
+    --extra-index-url https://download.pytorch.org/whl/cu126 \
+    "rmbg-as-a-service==0.0.2" hf-transfer
 
 EXPOSE 8000
 CMD ["rmbg-as-a-service"]

build/Makefile

@@ -0,0 +1,53 @@
+# mindthemath/rmbg image — build the published-package container, test it, push it.
+# Override inline, e.g.:  make release TAG=0.0.2
+
+COMPOSE ?= docker compose
+PYTHON  ?= python3
+IMAGE   ?= mindthemath/rmbg
+TAG     ?= latest
+PORT    ?= 8000
+INPUT   ?= ../test.jpg
+OUTPUT  ?= output.png
+BG      ?= alpha
+
+# Exported so compose.yml's ${TAG} interpolation picks it up.
+export TAG
+
+.DEFAULT_GOAL := help
+
+.PHONY: help build run up stop down logs log ps shell test push release clean
+
+help: ## Show this help
+	@grep -E '^[a-zA-Z_-]+:.*?## ' $(MAKEFILE_LIST) \
+		| awk 'BEGIN{FS=":.*?## "}{printf "  \033[36m%-12s\033[0m %s\n", $$1, $$2}'
+
+build: ## Build the mindthemath/rmbg image
+	$(COMPOSE) build
+
+run up: ## Start the service (GPU) in the background
+	$(COMPOSE) up -d
+
+stop down: ## Stop and remove the service container
+	$(COMPOSE) down
+
+logs log: ## Follow service logs
+	$(COMPOSE) logs -f
+
+ps: ## Show service status
+	$(COMPOSE) ps
+
+shell: ## Open a shell inside a fresh container
+	$(COMPOSE) run --rm --entrypoint bash rmbg
+
+test: ## Send INPUT to the running service, save OUTPUT
+	$(PYTHON) ../scripts/client.py --url http://localhost:$(PORT) \
+		--input $(INPUT) --output $(OUTPUT) --background $(BG)
+
+push: ## Push mindthemath/rmbg:$(TAG) to Docker Hub (needs docker login)
+	docker push $(IMAGE):$(TAG)
+
+release: build push ## Build then push mindthemath/rmbg:$(TAG)
+
+clean: ## Stop the service and remove the built image
+	-$(COMPOSE) down
+	-docker image rm $(IMAGE):$(TAG)

build/compose.yml

@@ -0,0 +1,31 @@
+# mindthemath/rmbg — runs the image built from the published (TestPyPI) package.
+services:
+  rmbg:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    image: mindthemath/rmbg:${TAG:-latest}
+    container_name: rmbg
+    ports:
+      - "${PORT:-8000}:8000"
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+      # Default variant/resolution; both are also selectable per request.
+      - BIREFNET_MODEL=${BIREFNET_MODEL:-general}
+      - BIREFNET_RESOLUTION=${BIREFNET_RESOLUTION:-1024}
+    # Use the nvidia-container-runtime for GPU acceleration.
+    runtime: nvidia
+    volumes:
+      # Persist downloaded BiRefNet weights across container restarts.
+      - hf-cache:/app/hf_cache
+    healthcheck:
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
+      interval: 15s
+      timeout: 5s
+      retries: 30
+      start_period: 180s
+    restart: unless-stopped
+
+volumes:
+  hf-cache: