Initial ComfyUI implementation with updated wrapper

workers/comfyui-json/README.md

@@ -0,0 +1,179 @@
+# ComfyUI PyWorker
+
+This is the base PyWorker for ComfyUI. It provides a unified interface for running any ComfyUI workflow through a proxy-based architecture.
+
+## Endpoint
+
+The worker provides a single endpoint:
+
+- `/generate/sync`: Processes ComfyUI workflows using either predefined modifiers or custom workflow JSON
+
+## Request Format
+
+The worker accepts requests in the following format. Choose either modifier mode OR custom workflow mode:
+
+**Modifier Mode:**
+```json
+{
+  "input": {
+    "request_id": "uuid-string",    // optional - UUID generated if not provided
+    "modifier": "RawWorkflow",
+    "modifications": {
+      "prompt": "a beautiful landscape",
+      "width": 1024,
+      "height": 1024,
+      "steps": 20,
+      "seed": 123456789
+    },
+    "s3": { ... },       // optional
+    "webhook": { ... }   // optional
+  },
+  "expected_time": 30.0
+}
+```
+
+**Custom Workflow Mode:**
+```json
+{
+  "input": {
+    "request_id": "uuid-string",    // optional - UUID generated if not provided
+    "workflow_json": {
+      // Complete ComfyUI workflow JSON
+    },
+    "s3": { ... },       // optional
+    "webhook": { ... }   // optional
+  },
+  "expected_time": 30.0
+}
+```
+
+## Request Fields
+
+### Required Fields
+
+- **`input`**: Contains the main workflow data
+- **`input.request_id`**: Unique identifier for the request
+- **`expected_time`**: Expected runtime in seconds on RTX4090 (defaults to 46.0 if not provided)
+
+### Workflow Mode (Choose One)
+
+You must provide either `modifier` OR `workflow_json`, but not both:
+
+#### Option 1: Modifier Mode
+- **`input.modifier`**: Name of the predefined workflow modifier (e.g., "Text2Image")
+- **`input.modifications`**: Parameters to pass to the modifier
+
+#### Option 2: Custom Workflow Mode  
+- **`input.workflow_json`**: Complete ComfyUI workflow JSON
+
+### Optional Fields
+
+- **`input.s3`**: S3 configuration for file storage
+- **`input.webhook`**: Webhook configuration for notifications
+
+These configurations can be provided in the request JSON or via environment variables. Request-level configuration takes precedence over environment variables.
+
+#### S3 Configuration
+
+**Via Request JSON:**
+```json
+"s3": {
+  "access_key_id": "your-s3-access-key",
+  "secret_access_key": "your-s3-secret-access-key", 
+  "endpoint_url": "https://my-endpoint.backblaze.com",
+  "bucket_name": "your-bucket",
+  "region": "us-east-1"
+}
+```
+
+**Via Environment Variables:**
+```bash
+S3_ACCESS_KEY_ID=your-key
+S3_SECRET_ACCESS_KEY=your-secret
+S3_BUCKET_NAME=your-bucket
+S3_ENDPOINT_URL=https://s3.amazonaws.com
+S3_REGION=us-east-1
+```
+
+#### Webhook Configuration
+
+**Via Request JSON:**
+```json
+"webhook": {
+  "url": "your-webhook-url",
+  "extra_params": {
+    "custom_field": "value"
+  }
+}
+```
+
+**Via Environment Variables:**
+```bash
+WEBHOOK_URL=https://your-webhook.com  # Default webhook URL
+WEBHOOK_TIMEOUT=30                   # Webhook timeout in seconds
+```
+
+## Examples
+
+### Basic Text-to-Image (Modifier Mode)
+
+```json
+{
+  "input": {
+    "modifier": "Text2Image",
+    "modifications": {
+      "prompt": "a cat sitting on a windowsill",
+      "width": 512,
+      "height": 512,
+      "steps": 20,
+      "seed": 42
+    }
+  },
+  "expected_time": 25.0
+}
+```
+
+### Custom Workflow Mode
+
+```json
+{
+  "input": {
+    "request_id": "67890",    // optional - using custom ID for tracking
+    "workflow_json": {
+      "3": {
+        "inputs": {
+          "seed": 42,
+          "steps": 20,
+          "cfg": 8,
+          "sampler_name": "euler",
+          "scheduler": "normal",
+          "denoise": 1,
+          "model": ["4", 0],
+          "positive": ["6", 0],
+          "negative": ["7", 0],
+          "latent_image": ["5", 0]
+        },
+        "class_type": "KSampler"
+      }
+    }
+  },
+  "expected_time": 45.0
+}
+```
+
+## Expected Time Guidelines
+
+The `expected_time` field helps with resource planning and should reflect expected runtime on RTX4090:
+
+- **Simple text-to-image**: 15-30 seconds
+- **Complex workflows with upscaling**: 60+ seconds  
+- **Video generation**: 180+ seconds
+- **Default**: 46 seconds (if not specified)
+
+## Client Libraries
+
+See the test client examples for implementation details on how to integrate with the ComfyUI worker.
+
+---
+
+See Vast's serverless documentation for more details on how to use ComfyUI with autoscaler.

workers/comfyui-json/client.py

@@ -0,0 +1,98 @@
+import logging
+import uuid
+import random
+from urllib.parse import urljoin
+
+import requests
+
+from lib.test_utils import print_truncate_res
+from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
+
+logging.basicConfig(
+    level=logging.DEBUG,
+    format="%(asctime)s[%(levelname)-5s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+log = logging.getLogger(__file__)
+
+
+def call_text2image_workflow(
+    endpoint_group_name: str, api_key: str, server_url: str
+) -> None:
+    """Simple Text2Image using the new modifier-based approach"""
+    WORKER_ENDPOINT = "/generate/sync"
+    COST = 100
+    
+    # Route to get worker URL
+    route_payload = {
+        "endpoint": endpoint_group_name,
+        "api_key": api_key,
+        "cost": COST,
+    }
+    response = requests.post(
+        urljoin(server_url, "/route/"),
+        json=route_payload,
+        timeout=4,
+    )
+    response.raise_for_status()
+    message = response.json()
+    url = message["url"]
+    auth_data = dict(
+        signature=message["signature"],
+        cost=message["cost"],
+        endpoint=message["endpoint"],
+        reqnum=message["reqnum"],
+        url=message["url"],
+    )
+    
+    # Build the new payload structure
+    payload = {
+        "input": {
+            "request_id": str(uuid.uuid4()),
+            "modifier": "RawWorkflow",  # or whatever your Text2Image modifier is called
+            "modifications": {
+                "prompt": "a beautiful landscape with mountains and lakes",
+                "width": 1024,
+                "height": 1024,
+                "steps": 20,
+                "seed": random.randint(0, 2**32 - 1)
+            },
+            "workflow_json": {}  # Empty since using modifier approach
+        },
+        "expected_time": 30.0  # Expected 30 seconds on RTX4090
+    }
+    
+    req_data = dict(payload=payload, auth_data=auth_data)
+    url = urljoin(url, WORKER_ENDPOINT)
+    print(f"url: {url}")
+    
+    response = requests.post(
+        url,
+        json=req_data,
+        verify=get_cert_file_path(),
+    )
+    response.raise_for_status()
+    print_truncate_res(str(response.json()))
+
+
+if __name__ == "__main__":
+    from lib.test_utils import test_args
+
+    args = test_args.parse_args()
+    endpoint_api_key = Endpoint.get_endpoint_api_key(
+        endpoint_name=args.endpoint_group_name,
+        account_api_key=args.api_key,
+        instance=args.instance,
+    )
+    if endpoint_api_key:
+        try:
+            call_text2image_workflow(
+                api_key=endpoint_api_key,
+                endpoint_group_name=args.endpoint_group_name,
+                server_url=args.server_url,
+            )
+        except Exception as e:
+            log.error(f"Error during API call: {e}")
+    else:
+        log.error(f"Failed to get API key for endpoint {args.endpoint_group_name}")

workers/comfyui-json/data_types.py

@@ -0,0 +1,82 @@
+import sys
+import json
+import random
+import dataclasses
+import inspect
+from typing import Dict, Any
+from functools import cache
+from math import ceil
+
+from lib.data_types import ApiPayload, JsonDataException
+
+
+with open("workers/comfyui/misc/test_prompts.txt", "r") as f:
+    test_prompts = f.readlines()
+
+@dataclasses.dataclass
+class ComfyWorkflowData(ApiPayload):
+    input: dict
+    expected_time: float = 46.0  # Default: 2x baseline (23s * 2) for RTX4090
+
+    @classmethod
+    def for_test(cls):
+        test_prompt = random.choice(test_prompts).rstrip()
+        return cls(
+            input={
+                "request_id": f"test-{random.randint(1000, 9999)}",
+                "modifier": "RawWorkflow",
+                "modifications": {
+                    "prompt": test_prompt,
+                    "width": 1024,
+                    "height": 1024,
+                    "steps": 28,
+                    "seed": random.randint(0, sys.maxsize),
+                }
+            },
+            expected_time=25.0  # Test data: expect 25 seconds on RTX4090 (slightly above baseline)
+        )
+
+    def generate_payload_json(self) -> Dict[str, Any]:
+        # input is already a dict, just return it wrapped in the expected structure
+        return {"input": self.input}
+
+    def count_workload(self) -> float:
+        """
+        This needs review. We cannot reasonably predict the workload based on the inputs. We may be processing:
+        - Images
+        - Videos
+        - Audio... There may also be complex loops in the workflow.
+
+        User will provide an expected time to complete and we will calculate equivalent cost
+
+        Convert user-provided expected_time (RTX4090 seconds) to the old scoring system.
+        
+        The old system normalized to: 1024x1024, 28 steps = 200 tokens on RTX4090
+        The old formula was: REQUEST_TIME_FOR_STANDARD_IMAGE * (time_ratio * 200)
+        
+        Now the user provides the expected request time directly.
+        Default expected_time is 46s (2x baseline) if not specified.
+        """
+        # Baseline: standard image (1024x1024, 28 steps) = 23s = 200 tokens on RTX4090
+        RTX4090_BASELINE_TIME = 23.0  # seconds for standard image on RTX4090
+        BASELINE_TOKENS = 200  # tokens for standard image
+        
+        # Calculate time ratio compared to baseline
+        time_ratio = self.expected_time / RTX4090_BASELINE_TIME
+        
+        # Return workload score: time_ratio * baseline tokens
+        return time_ratio * BASELINE_TOKENS
+
+    @classmethod
+    def from_json_msg(cls, json_msg: Dict[str, Any]) -> "ComfyWorkflowData":
+        # Extract required fields
+        if "input" not in json_msg:
+            raise JsonDataException({"input": "missing parameter"})
+        
+        # expected_time is optional, uses default if not provided
+        expected_time = json_msg.get("expected_time", 46.0)  # Default: 2x baseline
+        
+        return cls(
+            input=json_msg["input"],
+            expected_time=float(expected_time)
+        )

workers/comfyui-json/misc/test_prompts.txt

@@ -0,0 +1,34 @@
+cartoon character of a person with a hoodie , in style of cytus and deemo, ork, gold chains, realistic anime cat, dripping black goo, lineage revolution style, thug life, cute anthropomorphic bunny, balrog, arknights, aliased, very buff, black and red and yellow paint, painting illustration collage style, character composition in vector with white background
+stardew valley, fine details
+2D Vector Illustration of a child with soccer ball Art for Sublimation, Design Art, Chrome Art, Painting and Stunning Artwork, Highly Detailed Digital Painting, Airbrush Art, Highly Detailed Digital Artwork, Dramatic Artwork, stained antique yellow copper paint, digital airbrush art, detailed by Mark Brooks, Chicano airbrush art, Swagger! snake Culture
+realistic futuristic city-downtown with short buildings, sunset
+seascape by Ray Collins and artgerm, front view of a perfect wave, sunny background, ultra detailed water
+inspired by realflow-cinema4d editor features, create image of a transparent luxury cup with ice fruits and mint, connected with white, yellow and pink cream, Slow - High Speed MO Photography, YouTube Video Screenshot, Abstract Clay, Transparent Cup , molecular gastronomy, wheel, 3D fluid,Simulation rendering, still video, 4k polymer clay futras photography, very surreal, Houdini Fluid Simulation, hyperrealistic CGI and FLUIDS & MULTIPHYSICS SIMULATION effect, with Somali Stain Lurex, Metallic Jacquard, Gold Thread, Mulberry Silk, Toub Saree, Warm background, a fantastic image worthy of an award.
+biker with backpack on his back riding a motorcycle, Style by Ade Santora, Oilpunk, Cover photo, craig mullins style, on the cover of a magazine, Outdoor Magazine, inspired by Alex Petruk APe, image of a male biker, Cover of an award-winning magazine, the man has a backpack, photo for magazine, with a backpack, magazine cover
+generate a collage-style illustration inspired by the Procreate raster graphic editor, photographic illustration with the theme, 2D vector, art for textile sublimation, containing surrealistic cartoon cat wearing a baseball cap and jeans standing in front of a poster, inspired by Sadao Watanabe, Doraemon, Japanese cartoon style, Eichiro Oda, Iconic high detail character, Director: Nakahara Nantenbō, Kastuhiro Otomo, image detailed, by Miyamoto, Hidetaka Miyazaki, Katsuhiro illustration, 8k, masterpiece, Minimize noise and grain in photo quality without lose quality and increase brightness and lighting,Symmetry and Alignment, Avoid asymmetrical shapes and out-of-focus points. Focus and Sharpness: Make sure the image is focused and sharp and encourages the viewer to see it as a work of art printed on fabric.
+fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details
+Iron Man, (Arnold Tsang, Toru Nakayama), Masterpiece, Studio Quality, 6k , toa, toaair, 1boy, glowing, axe, mecha, science_fiction, solo, weapon, jungle , green_background, nature, outdoors, solo, tree, weapon, mask, dynamic lighting, detailed shading, digital texture painting
+(Pope Francis) wearing leather jacket is a DJ in a nightclub, mixing live on stage, giant mixing table, a masterpiece
+Pope Francis wearing biker (leather jacket), a masterpiece
+Luke Skywalker ordering a burger and fries from the Death Star canteen.
+I want to generate a group avatar for a Feishu group chat. The role of this group is daily software technical communication. Now the subject technology stacks that members of this group discuss daily include: algorithms, data structures, optimization, functional programming, and the programming languages often discussed are: TypeScript, Java, python, etc. I hope this avatar has a simple aesthetic, this avatar is a single person avatar
+portrait Anime black girl cute-fine-face, pretty face, realistic shaded Perfect face, fine details. Anime. realistic shaded lighting by Ilya Kuvshinov Giuseppe Dangelico Pino and Michael Garmash and Rob Rey, IAMAG premiere, WLOP matte print, cute freckles, masterpiece
+young Disney socialite wearing a beige miniskirt, dark brown turtleneck sweater, small neckless, cute-fine-face, anime. illustration, realistic shaded perfect face, brown hair, grey eyes, fine details, realistic shaded lighting by ilya kuvshinov giuseppe dangelico pino and michael garmash and rob rey, iamag premiere, wlop matte print, a masterpiece
+Cute small cat sitting in a movie theater eating chicken wiggs watching a movie ,unreal engine, cozy indoor lighting, artstation, detailed, digital painting,cinematic,character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render
+Cute small dog sitting in a movie theater eating popcorn watching a movie ,unreal engine, cozy indoor lighting, artstation, detailed, digital painting,cinematic,character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render
+fox bracelet made of buckskin with fox features, rich details, fine carvings, studio lighting
+crane buckskin bracelet with crane features, rich details, fine carvings, studio lighting
+london luxurious interior living-room, light walls
+Parisian luxurious interior penthouse bedroom, dark walls, wooden panels
+cute girl, crop-top, blond hair, black glasses, stretching, with background by greg rutkowski makoto shinkai kyoto animation key art feminine mid shot
+houses in front, houses background, straight houses, digital art, smooth, sharp focus, gravity falls style, doraemon style, shinchan style, anime style
+Simplified technical drawing, Leonardo da Vinci, Mechanical Dinosaur Skeleton, Minimalistic annotations, Hand-drawn illustrations, Basic design and engineering, Wonder and curiosity
+High quality 8K painting impressionist style of a Japanese modern city street with a girl on the foreground wearing a traditional wedding dress with a fox mask, staring at the sky, daylight
+a landscape from the Moon with the Earth setting on the horizon, realistic, detailed
+Isometric Atlantis city,great architecture with columns, great details, ornaments,seaweed, blue ambiance, 3D cartoon style, soft light, 45° view
+A hyper realistic avatar of a guy riding on a black honda cbr 650r in leather suit,high detail, high quality,8K,photo realism
+the street of amedieval fantasy town, at dawn, dark, highly detailed
+overwhelmingly beautiful eagle framed with vector flowers, long shiny wavy flowing hair, polished, ultra detailed vector floral illustration mixed with hyper realism, muted pastel colors, vector floral details in background, muted colors, hyper detailed ultra intricate overwhelming realism in detailed complex scene with magical fantasy atmosphere, no signature, no watermark
+a highly detailed matte painting of a man on a hill watching a rocket launch in the distance by studio ghibli, makoto shinkai, by artgerm, by wlop, by greg rutkowski, volumetric lighting, octane render, 4 k resolution, trending on artstation, masterpiece | hyperrealism| highly detailed| insanely detailed| intricate| cinematic lighting| depth of field
+electronik robot and ofice ,unreal engine, cozy indoor lighting, artstation, detailed, digital painting,cinematic,character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render
+exquisitely intricately detailed illustration, of a small world with a lake and a rainbow, inside a closed glass jar.

workers/comfyui-json/server.py

@@ -0,0 +1,116 @@
+import os
+import logging
+import dataclasses
+import base64
+from typing import Optional, Union, Type
+
+from aiohttp import web, ClientResponse
+
+from lib.backend import Backend, LogAction
+from lib.data_types import EndpointHandler
+from lib.server import start_server
+from .data_types import ComfyWorkflowData
+
+
+MODEL_SERVER_URL = "http://127.0.0.1:18288"
+
+# This is the last log line that gets emitted once comfyui+extensions have been fully loaded
+MODEL_SERVER_START_LOG_MSG = "To see the GUI go to: http://127.0.0.1:18188"
+MODEL_SERVER_ERROR_LOG_MSGS = [
+    "MetadataIncompleteBuffer",  # This error is emitted when the downloaded model is corrupted
+    "Value not in list: unet_name",  # This error is emitted when the model file is not there at all
+]
+
+
+logging.basicConfig(
+    level=logging.DEBUG,
+    format="%(asctime)s[%(levelname)-5s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+log = logging.getLogger(__file__)
+
+
+async def generate_client_response(
+        self, client_request: web.Request, model_response: ClientResponse
+    ) -> Union[web.Response, web.StreamResponse]:
+        # Check if the response is actually streaming based on response headers/content-type
+        is_streaming_response = (
+            model_response.content_type == "text/event-stream"
+            or model_response.content_type == "application/x-ndjson"
+            or model_response.headers.get("Transfer-Encoding") == "chunked"
+            or "stream" in model_response.content_type.lower()
+        )
+
+        if is_streaming_response:
+            log.debug("Detected streaming response...")
+            res = web.StreamResponse()
+            res.content_type = model_response.content_type
+            await res.prepare(client_request)
+            async for chunk in model_response.content:
+                await res.write(chunk)
+            await res.write_eof()
+            log.debug("Done streaming response")
+            return res
+        else:
+            log.debug("Detected non-streaming response...")
+            content = await model_response.read()
+            return web.Response(
+                body=content,
+                status=model_response.status,
+                content_type=model_response.content_type
+            )
+            
+
+@dataclasses.dataclass
+class ComfyWorkflowHandler(EndpointHandler[ComfyWorkflowData]):
+
+    @property
+    def endpoint(self) -> str:
+        return "/generate/sync"
+
+    @property
+    def healthcheck_endpoint(self) -> Optional[str]:
+        return None
+
+    @classmethod
+    def payload_cls(cls) -> Type[ComfyWorkflowData]:
+        return ComfyWorkflowData
+
+    def make_benchmark_payload(self) -> ComfyWorkflowData:
+        return ComfyWorkflowData.for_test()
+
+    async def generate_client_response(
+        self, client_request: web.Request, model_response: ClientResponse
+    ) -> Union[web.Response, web.StreamResponse]:
+        return await generate_client_response(client_request, model_response)
+
+
+backend = Backend(
+    model_server_url=MODEL_SERVER_URL,
+    model_log_file=os.environ["MODEL_LOG"],
+    allow_parallel_requests=False,
+    benchmark_handler=ComfyWorkflowHandler(
+        benchmark_runs=3, benchmark_words=100
+    ),
+    log_actions=[
+        (LogAction.ModelLoaded, MODEL_SERVER_START_LOG_MSG),
+        (LogAction.Info, "Downloading:"),
+        *[
+            (LogAction.ModelError, error_msg)
+            for error_msg in MODEL_SERVER_ERROR_LOG_MSGS
+        ],
+    ],
+)
+
+
+async def handle_ping(_):
+    return web.Response(body="pong")
+
+
+routes = [
+    web.post("/generate/sync", backend.create_handler(ComfyWorkflowHandler())),
+    web.get("/ping", handle_ping),
+]
+
+if __name__ == "__main__":
+    start_server(backend, routes)

workers/comfyui-json/test_load.py

@@ -0,0 +1,8 @@
+from lib.test_utils import test_load_cmd, test_args
+from .data_types import ComfyWorkflowData
+
+WORKER_ENDPOINT = "/generate/sync"
+
+
+if __name__ == "__main__":
+    test_load_cmd(ComfyWorkflowData, WORKER_ENDPOINT, arg_parser=test_args)