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",
+  "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 = "simonbaby/int8wo"
+revision = "ea08d478d1c800affec1dc0ea6442a6fa531bbb9"
+
+[[tool.edge-maxxing.models]]
+repository = "simonbaby/bf16"
+revision = "24a77356026c2b8552488a3381fef097ead3459d"
+
+[[tool.edge-maxxing.models]]
+repository = "simonbaby/vae_e3m2"
+revision = "3e2254259485390a5eb428ba82f98dcec24895f5"
+
+
+[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 +1,50 @@
+#3
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import torch._dynamo
+import gc
+import os
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from PIL import Image as img
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+
+Pipeline = None
+
+ids = "black-forest-labs/FLUX.1-schnell"
+Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+
+def load_pipeline() -> Pipeline:
+    text_encoder_2 = T5EncoderModel.from_pretrained("simonbaby/bf16", revision = "24a77356026c2b8552488a3381fef097ead3459d", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+    path = os.path.join(HF_HUB_CACHE, "models--simonbaby--int8wo/snapshots/ea08d478d1c800affec1dc0ea6442a6fa531bbb9")
+    model = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
+    pipeline = DiffusionPipeline.from_pretrained(ids, revision=Revision, transformer=model, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
+    pipeline.to("cuda")
+    quantize_(pipeline.vae, int8_weight_only())
+    for _ in range(3):
+        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) -> Image:
+    generator = Generator(pipeline.device).manual_seed(request.seed)
+
+    return pipeline(
+        request.prompt,
+        generator=generator,
+        guidance_scale=0.0,
+        num_inference_steps=4,
+        max_sequence_length=256,
+        height=request.height,
+        width=request.width,
+    ).images[0]