stage

.gitmodules

@@ -0,0 +1,4 @@
+[submodule "newdream-sdxl-20"]
+	path = models/newdream-sdxl-20
+	url = https://huggingface.co/stablediffusionapi/newdream-sdxl-20
+	branch = main

README.md

@@ -0,0 +1,12 @@
+# edge-maxxing-newdream-sdxl
+
+This holds the baseline for the SDXL Nvidia GeForce RTX 4090 contest, which can be forked freely and optimized
+
+Some recommendations are as follows:
+- Installing dependencies should be done in pyproject.toml, including git dependencies
+- Compiled models should be included directly in the repository(rather than compiling during loading), loading time matters far more than file sizes
+- Avoid changing `src/main.py`, as that includes mostly protocol logic. Most changes should be in `models` and `src/pipeline.py`
+- Change `install_args.txt` to add `pip install` arguments to be used when installing the package
+
+For testing, you need a docker container with pytorch and ubuntu 22.04,
+you can download your listed dependencies with `pip install $(cat install_args.txt) -e .`, and then running `start_inference`

install_args.txt

@@ -0,0 +1 @@
+

pyproject.toml

@@ -0,0 +1,19 @@
+[build-system]
+requires = ["setuptools >= 61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "edge-maxxing-4090-newdream"
+description = "An edge-maxxing model submission for the 4090 newdream contest"
+requires-python = ">=3.10,<3.11"
+version = "1.0.0"
+dependencies = [
+    "diffusers==0.30.2",
+    "transformers==4.41.2",
+    "accelerate==0.31.0",
+    "omegaconf==2.3.0",
+    "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing#subdirectory=pipelines",
+]
+
+[project.scripts]
+start_inference = "main:main"

src/main.py

@@ -0,0 +1,48 @@
+from io import BytesIO
+from multiprocessing.connection import Listener
+from os import chmod
+from os.path import abspath
+from pathlib import Path
+
+from PIL.JpegImagePlugin import JpegImageFile
+from pipelines.models import TextToImageRequest
+
+from pipeline import load_pipeline, infer
+
+SOCKET = abspath(Path(__file__).parent.parent / "inferences.sock")
+
+
+def main():
+    print(f"Loading pipeline")
+    pipeline = load_pipeline()
+
+    print(f"Pipeline loaded")
+
+    print(f"Creating socket at '{SOCKET}'")
+    with Listener(SOCKET) as listener:
+        chmod(SOCKET, 0o777)
+
+        print(f"Awaiting connections")
+        with listener.accept() as connection:
+            print(f"Connected")
+
+            while True:
+                try:
+                    request = TextToImageRequest.model_validate_json(connection.recv_bytes().decode("utf-8"))
+                except EOFError:
+                    print(f"Inference socket exiting")
+
+                    return
+
+                image = infer(request, pipeline)
+
+                data = BytesIO()
+                image.save(data, format=JpegImageFile.format)
+
+                packet = data.getvalue()
+
+                connection.send_bytes(packet)
+
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,29 @@
+import torch
+from PIL.Image import Image
+from diffusers import StableDiffusionXLPipeline
+from pipelines.models import TextToImageRequest
+from torch import Generator
+
+
+def load_pipeline() -> StableDiffusionXLPipeline:
+    pipeline = StableDiffusionXLPipeline.from_pretrained(
+        "./models/newdream-sdxl-20",
+        torch_dtype=torch.float16,
+        local_files_only=True,
+    ).to("cuda")
+
+    pipeline(prompt="")
+
+    return pipeline
+
+
+def infer(request: TextToImageRequest, pipeline: StableDiffusionXLPipeline) -> Image:
+    generator = Generator(pipeline.device).manual_seed(request.seed) if request.seed else None
+
+    return pipeline(
+        prompt=request.prompt,
+        negative_prompt=request.negative_prompt,
+        width=request.width,
+        height=request.height,
+        generator=generator,
+    ).images[0]