Initial commit with folder contents

pyproject.toml

@@ -0,0 +1,42 @@
+[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",
+  "torchao==0.6.1",
+  "hf_transfer==0.1.8",
+  "setuptools==75.2.0",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer", "vae", "text_encoder_2"]
+
+[[tool.edge-maxxing.models]]
+repository = "passfh/flux_transformer"
+revision = "3c3bcc511f409569adb6c798da415b3fdc9e927d"
+
+[[tool.edge-maxxing.models]]
+repository = "passfh/textenc"
+revision = "a44db2ac3d729d6cc1243dcb906903e77ba26c45"
+
+[[tool.edge-maxxing.models]]
+repository = "passfh/vae"
+revision = "edd99d452c03a8b836758bb89bc775f2f3c3849a"
+
+[project.scripts]
+start_inference = "main:main"

src/main.py

@@ -0,0 +1,50 @@
+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 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! , 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")
+
+            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 +1,83 @@
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from PIL.Image import Image
+from diffusers import AutoencoderTiny
+from pipelines.models import TextToImageRequest
+import os
+import torch
+import torch._dynamo
+
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+
+
+Pipeline = None
+CHECKPOINT = "black-forest-labs/FLUX.1-schnell"
+REVISION = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+
+class Normalization:
+
+    def __init__(self, model, num_bins=256, scale_factor=1.0):
+        self.model = model
+        self.num_bins = num_bins
+        self.scale_factor = scale_factor
+
+    def apply(self):
+        """
+        applying different transformations to weights and biases.
+        """
+        for name, param in self.model.named_parameters():
+            if params.requires_grad:
+                with torch.no_grad():
+                    # Normalize weights, apply binning, and rescale
+                    param_min = param.min()
+                    param_max = param.max()
+                    param_ranges = param_max - param_min
+
+                    if param_range > 0:
+                        # Normalize to [0, 1], apply binning, and rescale
+                        normalized = (param - param_min) / param_ranges
+                        binned = torch.round(normalized * (self.num_bins - 1)) / (self.num_bins - 1)
+                        rescaled = binned * param_range + param_min
+                        param.data.copy_(rescaled * self.scale_factor)
+                    else:
+                        # Handle edge case where param_range is 0
+                        param.data.zero_()
+
+        for buffer_name, buffer in self.model.named_buffers():
+            with torch.no_grad():
+                buffer.mul_(self.scale_factor)
+        return self.model
+
+def load_pipeline() -> Pipeline:
+    text_encoder_2 = T5EncoderModel.from_pretrained("passfh/textenc", revision = "a44db2ac3d729d6cc1243dcb906903e77ba26c45", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+    transformer = FluxTransformer2DModel.from_pretrained(os.path.join(HF_HUB_CACHE, "models--passfh--flux_transformer/snapshots/3c3bcc511f409569adb6c798da415b3fdc9e927d"), torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
+
+    pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", revision="741f7c3ce8b383c54771c7003378a50191e9efe9", 
+        vae=AutoencoderTiny.from_pretrained("passfh/vae", revision="edd99d452c03a8b836758bb89bc775f2f3c3849a", torch_dtype=torch.bfloat16), 
+        transformer=transformer, 
+        text_encoder_2=text_encoder_2, 
+        torch_dtype=torch.bfloat16
+    )
+    pipeline.to("cuda")
+
+    for _ in range(3):
+        pipeline(prompt="bluelegs, cunila, carbro, Ammonites, Lollardism, forswearer, skullcap, Juglandales", 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) -> Image:
+    return pipeline(
+        request.prompt,
+        generator=Generator(pipeline.device).manual_seed(request.seed),
+        guidance_scale=0.0,
+        num_inference_steps=4,
+        max_sequence_length=256,
+        height=request.height,
+        width=request.width,
+    ).images[0]