feat(inference): Add SD CPP module

.env.sample

@@ -6,3 +6,12 @@ POSTGRES_PORT=your_db_port
 KC_DB_SCHEMA=public
 KEYCLOAK_ADMIN=your_keycloak_admin_username
 KEYCLOAK_ADMIN_PASSWORD=your_keycloak_admin_password
+
+# Inference
+## LLM
+MODEL_URL=https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/resolve/main/llama-2-7b-chat.ggmlv3.q4_1.bin
+LLM_MODEL_FILE=$(basename $MODEL_URL)
+
+## SD
+MODEL_URL=https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
+SD_MODEL_FILE=$(basename $MODEL_URL)

.gitignore

@@ -1,2 +1,11 @@
 .vscode
 .env
+
+# Jan inference
+jan-inference/llm/models/**
+jan-inference/llm/.env
+
+jan-inference/sd/models/**
+jan-inference/sd/output/**
+jan-inference/sd/.env
+jan-inference/sd/sd

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "jan-inference/sd/sd_cpp"]
+	path = jan-inference/sd/sd_cpp
+	url = https://github.com/leejet/stable-diffusion.cpp

jan-inference/llm/README.md

@@ -0,0 +1,8 @@
+# Inference - LLM
+
+```bash
+docker network create traefik_public
+cp .env.example .env
+# -> Update MODEL_URL in `.env` file
+docker compose up -d --scale llm=2
+``````

jan-inference/llm/docker-compose.yml

@@ -0,0 +1,94 @@
+version: '3'
+
+services:
+
+  # Service to download a model file.
+  downloader:
+    image: busybox
+    # The command extracts the model filename from MODEL_URL and downloads it if it doesn't exist.
+    command: /bin/sh -c "LLM_MODEL_FILE=$(basename ${MODEL_URL}); if [ ! -f /models/$LLM_MODEL_FILE ]; then wget -O /models/$LLM_MODEL_FILE ${MODEL_URL}; fi"
+    # Mount a local directory to store the downloaded model.
+    volumes:
+      - ./models:/models
+
+  # Service to wait for the downloader service to finish downloading the model.
+  wait-for-downloader:
+    image: busybox
+    # The command waits until the model file (specified in MODEL_URL) exists.
+    command: /bin/sh -c "LLM_MODEL_FILE=$(basename ${MODEL_URL}); echo 'Waiting for downloader to finish'; while [ ! -f /models/$LLM_MODEL_FILE ]; do sleep 1; done; echo 'Model downloaded!'"
+    # Specifies that this service should start after the downloader service has started.
+    depends_on:
+      downloader:
+        condition: service_started
+    # Mount the same local directory to check for the downloaded model.
+    volumes:
+      - ./models:/models
+
+  # Service to run the Llama web application.
+  llm:
+    image: ghcr.io/abetlen/llama-cpp-python:latest
+    # Mount the directory that contains the downloaded model.
+    volumes:
+      - ./models:/models
+    environment:
+      # Specify the path to the model for the web application.
+      MODEL: /models/llama-2-7b-chat.ggmlv3.q4_1.bin
+    labels:
+      # Instead of using the Host rule, set a PathPrefix rule
+      - "traefik.http.routers.web.rule=PathPrefix(`/inference/llm`)"
+      # This tells Traefik where to forward the traffic for this route.
+      - "traefik.http.routers.web.service=llm"
+    # Define a service for the llm and specify its load balancer configuration
+      - "traefik.http.services.llm-service.loadbalancer.server.port=8000"
+
+      - "traefik.http.middlewares.strip-llm-prefix.stripprefix.prefixes=/inference/llm"
+      - "traefik.http.routers.web.middlewares=strip-llm-prefix"
+    # Health check configuration
+    healthcheck:
+      test: ["CMD", "wget", "--quiet", "--tries=1", "--spider", "http://localhost:8000"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 30s
+    # Restart policy configuration
+    restart: on-failure
+    # Specifies that this service should start only after wait-for-downloader has completed successfully.
+    depends_on:
+      wait-for-downloader:
+        condition: service_completed_successfully
+    # Connect this service to two networks: inference_net and traefik_public.
+    networks:
+      - inference_net
+      - traefik_public
+  
+  # Service for Traefik, a modern HTTP reverse proxy and load balancer.
+  traefik:
+    image: traefik:v2.5
+    command:
+      # Enable the Traefik API dashboard without TLS (not recommended for production).
+      - "--api.insecure=true"
+      # Enable Traefik to use Docker as a provider.
+      - "--providers.docker=true"
+      # Do not expose services by default. Explicitly specify in each service if it should be exposed.
+      - "--providers.docker.exposedbydefault=false"
+      # Specify the default entry point on port 80.
+      - "--entrypoints.web.address=:80"
+    ports:
+      # Map port 80 in the container to port 80 on the host.
+      - "80:80"
+      # Map port 8080 in the container (Traefik's dashboard) to port 8080 on the host.
+      - "8080:8080"
+    # Mount the Docker socket to allow Traefik to listen to Docker's API.
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock
+    # Connect this service to the traefik_public network.
+    networks:
+      - traefik_public
+
+# Define networks used in this docker-compose file.
+networks:
+  # Network for the llm service (used for inference).
+  inference_net:
+  # Public-facing network that Traefik uses. Marked as external to indicate it may be defined outside this file.
+  traefik_public:
+    external: true

jan-inference/sd/.dockerignore

@@ -0,0 +1,7 @@
+output/
+models/
+
+sd_cpp/.git
+sd_cpp/.github
+
+sd

jan-inference/sd/compile.Dockerfile

@@ -0,0 +1,10 @@
+FROM python:3.9.17
+
+RUN curl https://sh.rustup.rs -sSf | bash -s -- -y
+ENV PATH=/root/.cargo/bin:$PATH
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN pip install -r compile.requirements.txt

jan-inference/sd/compile.requirements.txt

@@ -0,0 +1,17 @@
+# Compiling
+-r sd_cpp/models/requirements.txt
+
+# diffusers
+# torch
+# ftfy
+# scipy
+# transformers
+# accelerate
+# huggingface-hub
+# xformers
+# omegaconf
+# safetensors
+# cog
+# tomesd
+# compel
+

jan-inference/sd/docker-compose.yml

@@ -0,0 +1,98 @@
+version: '3'
+
+services:
+
+  # Service to download a model file.
+  downloader:
+    build:
+      context: ./
+      dockerfile: compile.Dockerfile
+    # platform: "linux/amd64"
+    # The command extracts the model filename from MODEL_URL and downloads it if it doesn't exist.
+    command: /bin/sh -c "SD_MODEL_FILE=$(basename ${MODEL_URL}); if [ ! -f /converted_models/*.bin ]; then wget -O /converted_models/$SD_MODEL_FILE ${MODEL_URL}; python /sd.cpp/models/convert.py --out_type q4_0 --out_file /converted_models/$SD_MODEL_FILE; fi"
+    # Mount a local directory to store the downloaded model.
+    volumes:
+      - ./models:/converted_models
+
+  # Service to wait for the downloader service to finish downloading the model.
+  wait-for-downloader:
+    image: busybox
+    # The command waits until the model file (specified in MODEL_URL) exists.
+    command: /bin/sh -c "SD_MODEL_FILE=$(basename ${MODEL_URL}); echo 'Waiting for downloader to finish'; while [ ! -f /models/*.bin ]; do sleep 1; done; echo 'Model downloaded and converted!'"
+    # Specifies that this service should start after the downloader service has started.
+    depends_on:
+      downloader:
+        condition: service_started
+    # Mount the same local directory to check for the downloaded model.
+    volumes:
+      - ./models:/models
+
+  # Service to run the SD web application.
+  sd:
+    build:
+      context: ./
+      dockerfile: inference.Dockerfile
+    # Mount the directory that contains the downloaded model.
+    volumes:
+      - ./models:/models
+      - ./output/:/serving/output
+    command:  /bin/bash -c "python -m uvicorn main:app --proxy-headers --host 0.0.0.0 --port 8000"
+    # platform: "linux/amd64"
+    environment:
+      # Specify the path to the model for the web application.
+      BASE_URL: http://0.0.0.0:8000
+      MODEL_NAME: "v1-5-pruned-emaonly-ggml-model-q5_0.bin"
+      MODEL_DIR: "/models"
+      SD_PATH: "/sd"
+      PYTHONUNBUFFERED: 1
+    ports:
+      - 8000:8000
+    # Health check configuration
+    healthcheck:
+      test: ["CMD", "wget", "--quiet", "--tries=1", "--spider", "http://localhost:8000"]
+      interval: 30s
+      timeout: 10s
+      retries: 3
+      start_period: 30s
+    # Restart policy configuration
+    restart: on-failure
+    # Specifies that this service should start only after wait-for-downloader has completed successfully.
+    depends_on:
+      wait-for-downloader:
+        condition: service_completed_successfully
+    # Connect this service to two networks: inference_net and traefik_public.
+    networks:
+      - inference_net
+      - traefik_public
+  
+  # Service for Traefik, a modern HTTP reverse proxy and load balancer.
+  traefik:
+    image: traefik:v2.5
+    command:
+      # Enable the Traefik API dashboard without TLS (not recommended for production).
+      - "--api.insecure=true"
+      # Enable Traefik to use Docker as a provider.
+      - "--providers.docker=true"
+      # Do not expose services by default. Explicitly specify in each service if it should be exposed.
+      - "--providers.docker.exposedbydefault=false"
+      # Specify the default entry point on port 80.
+      - "--entrypoints.web.address=:80"
+    ports:
+      # Map port 80 in the container to port 80 on the host.
+      - "80:80"
+      # Map port 8080 in the container (Traefik's dashboard) to port 8080 on the host.
+      - "8080:8080"
+    # Mount the Docker socket to allow Traefik to listen to Docker's API.
+    volumes:
+      - /var/run/docker.sock:/var/run/docker.sock
+    # Connect this service to the traefik_public network.
+    networks:
+      - traefik_public
+
+# Define networks used in this docker-compose file.
+networks:
+  # Network for the llm service (used for inference).
+  inference_net:
+  # Public-facing network that Traefik uses. Marked as external to indicate it may be defined outside this file.
+  traefik_public:
+    external: true

jan-inference/sd/inference.Dockerfile

@@ -0,0 +1,19 @@
+FROM python:3.9.17 as build
+
+RUN apt-get update && apt-get install -y build-essential git cmake
+
+WORKDIR /sd.cpp
+
+COPY sd_cpp /sd.cpp
+
+RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+
+FROM python:3.9.17 as runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+WORKDIR /serving
+
+COPY . /serving/
+
+RUN pip install -r inference.requirements.txt

jan-inference/sd/inference.requirements.txt

@@ -0,0 +1,4 @@
+# Inference
+fastapi
+uvicorn
+python-multipart

jan-inference/sd/main.py

@@ -0,0 +1,70 @@
+from fastapi import FastAPI, BackgroundTasks, HTTPException, Form
+from fastapi.responses import FileResponse
+from fastapi.staticfiles import StaticFiles
+import subprocess
+import os
+from uuid import uuid4
+
+app = FastAPI()
+
+OUTPUT_DIR = "output"
+SD_PATH = os.environ.get("SD_PATH", "./sd")
+MODEL_DIR = os.environ.get("MODEL_DIR", "./models")
+BASE_URL = os.environ.get("BASE_URL", "http://localhost:8000")
+MODEL_NAME = os.environ.get(
+    "MODEL_NAME", "v1-5-pruned-emaonly-ggml-model-q5_0.bin")
+
+# Create the OUTPUT_DIR directory if it does not exist
+if not os.path.exists(OUTPUT_DIR):
+    os.makedirs(OUTPUT_DIR)
+
+# Create the OUTPUT_DIR directory if it does not exist
+if not os.path.exists(MODEL_DIR):
+    os.makedirs(MODEL_DIR)
+
+# Serve files from the "files" directory
+app.mount("/output", StaticFiles(directory=OUTPUT_DIR), name="output")
+
+
+def run_command(prompt: str, filename: str):
+    # Construct the command based on your provided example
+    command = [SD_PATH,
+               "-m", os.path.join(MODEL_DIR, MODEL_NAME),
+               "-p", prompt,
+               "-o", os.path.join(OUTPUT_DIR, filename)
+               ]
+
+    try:
+        sub_output = subprocess.run(command, timeout=5*60, capture_output=True,
+                                    check=True, encoding="utf-8")
+        print(sub_output.stdout)
+    except subprocess.CalledProcessError:
+        raise HTTPException(
+            status_code=500, detail="Failed to execute the command.")
+
+
+@app.post("/inference/")
+async def run_inference(background_tasks: BackgroundTasks, prompt: str = Form()):
+    # Generate a unique filename using uuid4()
+    filename = f"{uuid4()}.png"
+
+    # We will use background task to run the command so it won't block
+    background_tasks.add_task(run_command, prompt, filename)
+
+    # Return the expected path of the output file
+    return {"url": f'{BASE_URL}/serve/{filename}'}
+
+
+@app.get("/serve/{filename}")
+async def serve_file(filename: str):
+    file_path = os.path.join(OUTPUT_DIR, filename)
+
+    if os.path.exists(file_path):
+        return FileResponse(file_path)
+    else:
+        raise HTTPException(status_code=404, detail="File not found")
+
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=8000)

jan-inference/sd/sd_cpp

@@ -0,0 +1 @@
+Subproject commit 0d7f04b135cd48e8d62aecd09a52eb2afa482744