Initial commit with folder contents

.gitattributes

@@ -32,4 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.xz filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

pyproject.toml

@@ -0,0 +1,40 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "An edge-maxxing model submission for the 4090 Flux contest"
+requires-python = ">=3.10,<3.13"
+version = "8"
+dependencies = [
+  "diffusers==0.31.0",
+  "transformers==4.46.2",
+  "accelerate==1.1.0",
+  "omegaconf==2.3.0",
+  "torch==2.5.1",
+  "protobuf==5.28.3",
+  "sentencepiece==0.2.0",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+  "gitpython>=3.1.43",
+  "hf_transfer==0.1.8",
+  "torchao==0.6.1",
+  "setuptools>=75.3.0",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer"]
+
+[[tool.edge-maxxing.models]]
+repository = "db900/encode-stream"
+revision = "8a6b7bd09dc66733fa582900186f929353f63619"
+
+[[tool.edge-maxxing.models]]
+repository = "db900/deval"
+revision = "cf3aa8b38736807b622d7397036564a1cdc868de" 
+
+[project.scripts]
+start_inference = "main:main"
+

src/main.py

@@ -0,0 +1,55 @@
+import atexit
+from io import BytesIO
+from multiprocessing.connection import Listener
+from os import chmod, remove
+from os.path import abspath, exists
+from pathlib import Path
+from git import Repo
+import torch
+
+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 at_exit():
+    torch.cuda.empty_cache()
+
+
+def main():
+    atexit.register(at_exit)
+
+    print(f"Loading pipeline")
+    pipeline = load_pipeline()
+
+    print(f"Pipeline loaded! , creating socket at '{SOCKET}'")
+
+    if exists(SOCKET):
+        remove(SOCKET)
+
+    with Listener(SOCKET) as listener:
+        chmod(SOCKET, 0o777)
+
+        print(f"Awaiting connections")
+        with listener.accept() as connection:
+            print(f"Connected")
+            generator = torch.Generator("cuda")
+            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, generator.manual_seed(request.seed))
+                data = BytesIO()
+                image.save(data, format=JpegImageFile.format)
+
+                packet = data.getvalue()
+
+                connection.send_bytes(packet )
+
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,45 @@
+import os
+import gc
+import torch
+import torch._dynamo
+from PIL import Image
+from typing import Type
+from torch import Generator
+from transformers import T5EncoderModel
+from huggingface_hub.constants import HF_HUB_CACHE
+from pipelines.models import TextToImageRequest
+from torchao.quantization import quantize_, int8_weight_only
+from diffusers import AutoencoderKL, FluxPipeline, FluxTransformer2DModel
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.enabled = True
+torch.backends.cudnn.benchmark = True
+
+Pipeline = None
+
+def empty_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def load_pipeline() -> Pipeline:
+    ckpt_id = "black-forest-labs/FLUX.1-schnell"
+    ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+    
+    text_encoder_2 = T5EncoderModel.from_pretrained("db900/encode-stream", revision = "8a6b7bd09dc66733fa582900186f929353f63619", subfolder="text_encoder_2", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+    path = os.path.join(HF_HUB_CACHE, "models--db900--encode-stream/snapshots/8a6b7bd09dc66733fa582900186f929353f63619/transformer")
+    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
+    quantize_(AutoencoderKL.from_pretrained(ckpt_id, revision=ckpt_revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,), int8_weight_only())
+    pipeline = FluxPipeline.from_pretrained(ckpt_id, revision=ckpt_revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
+    pipeline.to("cuda")
+    with torch.inference_mode():
+        pipeline(prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    return pipeline
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
+    return pipeline(request.prompt, generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]