1movmedia/pyworker
2
0
Fork 0
Code Pull Requests Activity

Use cost value 1 for all jobs

Browse Source
This commit is contained in:
Rob Ballantyne
2025-08-25 17:58:22 +01:00
parent 92ff412679
commit ba74ac8136
2 changed files with 42 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
workers/comfyui-json/README.md
+26 -18
View File
@@ -2,6 +2,28 @@
This is the base PyWorker for ComfyUI. It provides a unified interface for running any ComfyUI workflow through a proxy-based architecture.
The cost for each request has a static value of `1`. ComfyUI does not handle concurrent workloads and there is no current provision to load multiple instances of ComfyUI per worker node.
## Requirements
This worker requires both [ComfyUI](https://github.com/comfyanonymous/ComfyUI) and [ComfyUI API Wrapper](https://github.com/ai-dock/comfyui-api-wrapper).
A docker image is provided but you may use any if the above requirements are met.
## Benchmarking
A simple image generation benchmark is run when the worker is first loaded.
This benchmark uses SD v1.5 in the default text to image workflow provided by ComfyUI. The following variables can be used to alter the complexity and running time of the benchmark:
| Environment Variable | Default Value |
| -------------------- | ------------- |
| BENCHMARK_TEST_WIDTH | 512 |
| BENCHMARK_TEST_HEIGHT | 512 |
| BENCHMARK_TEST_STEPS | 20 |
The prompt will be randomly selected from the file in misc/test_prompts.txt and a random seed used for every run of the benchmark.
## Endpoint
The worker provides a single endpoint:
@@ -27,8 +49,7 @@ The worker accepts requests in the following format. Choose either modifier mode
},
"s3": { ... }, // optional
"webhook": { ... } // optional
},
"expected_time": 30.0
}
}
```
@@ -42,8 +63,7 @@ The worker accepts requests in the following format. Choose either modifier mode
},
"s3": { ... }, // optional
"webhook": { ... } // optional
},
"expected_time": 30.0
}
}
```
@@ -53,7 +73,6 @@ The worker accepts requests in the following format. Choose either modifier mode
- **`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)
@@ -128,8 +147,7 @@ WEBHOOK_TIMEOUT=30 # Webhook timeout in seconds
"steps": 20,
"seed": 42
}
},
"expected_time": 25.0
}
}
```
@@ -156,20 +174,10 @@ WEBHOOK_TIMEOUT=30 # Webhook timeout in seconds
"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.
workers/comfyui-json/data_types.py
+16 -39
View File
@@ -1,8 +1,7 @@
import os
import sys
import json
import random
import dataclasses
import inspect
from typing import Dict, Any
from functools import cache
from math import ceil
@@ -13,27 +12,33 @@ from lib.data_types import ApiPayload, JsonDataException
with open("workers/comfyui/misc/test_prompts.txt", "r") as f:
test_prompts = f.readlines()
def count_workload() -> float:
# Always 1.0 where there is a single instance of ComfyUI handling requests
return 1.0
@dataclasses.dataclass
class ComfyWorkflowData(ApiPayload):
input: dict
expected_time: float = 46.0 # Default: 2x baseline (23s * 2) for RTX4090
@classmethod
def for_test(cls):
"""
Use the variables available to simulate workflows of the required running time
Example: SD1.5, simple image gen 10000 steps, 512px x 512px will run for approximately 9 minutes @ ~18 it/s (RTX 4090)
"""
test_prompt = random.choice(test_prompts).rstrip()
return cls(
input={
"request_id": f"test-{random.randint(1000, 9999)}",
"request_id": f"test-{random.randint(1000, 99999)}",
"modifier": "Text2Image",
"modifications": {
"prompt": test_prompt,
"width": 1024,
"height": 1024,
"steps": 28,
"width": os.getenv('BENCHMARK_TEST_WIDTH', 512),
"height": os.getenv('BENCHMARK_TEST_HEIGHT', 512),
"steps": os.getenv('BENCHMARK_TEST_STEPS', 20),
"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]:
@@ -41,31 +46,7 @@ class ComfyWorkflowData(ApiPayload):
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
return count_workload()
@classmethod
def from_json_msg(cls, json_msg: Dict[str, Any]) -> "ComfyWorkflowData":
@@ -73,10 +54,6 @@ class ComfyWorkflowData(ApiPayload):
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)
input=json_msg["input"]
)