Run Flux on ComfyUI as an API

modal serve 06_gpu_and_ml/comfyui/comfyapp.py

python 06_gpu_and_ml/comfyui/comfyclient.py --dev --modal-workspace $(modal profile current) --prompt "Surreal dreamscape with floating islands, upside-down waterfalls, and impossible geometric structures, all bathed in a soft, ethereal light"

import json
import subprocess
import uuid
from pathlib import Path
from typing import Dict

import modal

image = (  # build up a Modal Image to run ComfyUI, step by step
    modal.Image.debian_slim(  # start from basic Linux with Python
        python_version="3.11"
    )
    .apt_install("git")  # install git to clone ComfyUI
    .pip_install("fastapi[standard]==0.115.4")  # install web dependencies
    .pip_install("comfy-cli==1.3.5")  # install comfy-cli
    .run_commands(  # use comfy-cli to install ComfyUI and its dependencies
        "comfy --skip-prompt install --nvidia --version 0.3.10"
    )
)

image = (
    image.run_commands(  # download a custom node
        "comfy node install was-node-suite-comfyui@1.0.2"
    )
    # Add .run_commands(...) calls for any other custom nodes you want to download
)

def hf_download():
    from huggingface_hub import hf_hub_download

    flux_model = hf_hub_download(
        repo_id="Comfy-Org/flux1-schnell",
        filename="flux1-schnell-fp8.safetensors",
        cache_dir="/cache",
    )

    # symlink the model to the right ComfyUI directory
    subprocess.run(
        f"ln -s {flux_model} /root/comfy/ComfyUI/models/checkpoints/flux1-schnell-fp8.safetensors",
        shell=True,
        check=True,
    )


vol = modal.Volume.from_name("hf-hub-cache", create_if_missing=True)

image = (
    # install huggingface_hub with hf_transfer support to speed up downloads
    image.pip_install("huggingface_hub[hf_transfer]==0.26.2")
    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
    .run_function(
        hf_download,
        # persist the HF cache to a Modal Volume so future runs don't re-download models
        volumes={"/cache": vol},
    )
)

image = image.add_local_file(
    Path(__file__).parent / "workflow_api.json", "/root/workflow_api.json"
)

app = modal.App(name="example-comfyui", image=image)


@app.function(
    allow_concurrent_inputs=10,  # required for UI startup process which runs several API calls concurrently
    max_containers=1,  # limit interactive session to 1 container
    gpu="L40S",  # good starter GPU for inference
    volumes={"/cache": vol},  # mounts our cached models
)
@modal.web_server(8000, startup_timeout=60)
def ui():
    subprocess.Popen("comfy launch -- --listen 0.0.0.0 --port 8000", shell=True)

@app.cls(
    allow_concurrent_inputs=10,  # allow 10 concurrent API calls
    scaledown_window=300,  # 5 minute container keep alive after it processes an input
    gpu="L40S",
    volumes={"/cache": vol},
)
class ComfyUI:
    @modal.enter()
    def launch_comfy_background(self):
        # starts the ComfyUI server in the background exactly once when the first input is received
        cmd = "comfy launch --background"
        subprocess.run(cmd, shell=True, check=True)

    @modal.method()
    def infer(self, workflow_path: str = "/root/workflow_api.json"):
        # checks if ComfyUI server is healthy and accepting requests
        self.poll_server_health()

        # runs the comfy run --workflow command as a subprocess
        cmd = f"comfy run --workflow {workflow_path} --wait --timeout 1200 --verbose"
        subprocess.run(cmd, shell=True, check=True)

        # completed workflows write output images to this directory
        output_dir = "/root/comfy/ComfyUI/output"

        # looks up the name of the output image file based on the workflow
        workflow = json.loads(Path(workflow_path).read_text())
        file_prefix = [
            node.get("inputs")
            for node in workflow.values()
            if node.get("class_type") == "SaveImage"
        ][0]["filename_prefix"]

        # returns the image as bytes
        for f in Path(output_dir).iterdir():
            if f.name.startswith(file_prefix):
                return f.read_bytes()

    @modal.fastapi_endpoint(method="POST")
    def api(self, item: Dict):
        from fastapi import Response

        workflow_data = json.loads(
            (Path(__file__).parent / "workflow_api.json").read_text()
        )

        # insert the prompt
        workflow_data["6"]["inputs"]["text"] = item["prompt"]

        # give the output image a unique id per client request
        client_id = uuid.uuid4().hex
        workflow_data["9"]["inputs"]["filename_prefix"] = client_id

        # save this updated workflow to a new file
        new_workflow_file = f"{client_id}.json"
        json.dump(workflow_data, Path(new_workflow_file).open("w"))

        # run inference on the currently running container
        img_bytes = self.infer.local(new_workflow_file)

        return Response(img_bytes, media_type="image/jpeg")

    def poll_server_health(self) -> Dict:
        import socket
        import urllib

        try:
            # dummy request to check if the server is healthy
            req = urllib.request.Request("http://127.0.0.1:8188/system_stats")
            urllib.request.urlopen(req, timeout=5)
            print("ComfyUI server is healthy")
        except (socket.timeout, urllib.error.URLError) as e:
            # if no response in 5 seconds, stop the container; Modal will schedule queued inputs on a new container
            print(f"Server health check failed: {str(e)}")
            modal.experimental.stop_fetching_inputs()

            raise Exception("ComfyUI server is not healthy, stopping container")