Initial commit with folder contents

pyproject.toml

@@ -0,0 +1,29 @@
+[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 = "7"
+dependencies = [
+  "diffusers==0.31.0",
+  "transformers==4.46.2",
+  "accelerate==1.1.0",
+  "setuptools >= 75.0",
+  "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",
+  "torchao>=0.6.1",
+  "ipython>=8.29.0",
+  "gitpython>=3.1.43",
+]
+
+[tool.edge-maxxing]
+models = ["farapart/NEW_FLUX"]
+
+[project.scripts]
+start_inference = "main:main"

src/main.py

@@ -0,0 +1,59 @@
+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
+
+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")
+
+            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 +1,64 @@
+import torch
+from PIL import Image
+from diffusers import FluxPipeline
+from pipelines.models import TextToImageRequest
+from torch import Generator
+import os
+from diffusers import FluxPipeline, AutoencoderKL
+from diffusers.image_processor import VaeImageProcessor
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import diffusers
+import gc
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+
+from torchao.quantization import quantize_, int8_weight_only
+
+torch.set_float32_matmul_precision("high")
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.cuda.set_per_process_memory_fraction(0.99)
+
+
+Pipeline = None
+ckpt_id = "farapart/NEW_FLUX"
+
+
+def fart():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+
+def load_pipeline() -> Pipeline:
+
+    fart()
+
+    vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=torch.bfloat16)
+    quantize_(vae, int8_weight_only())
+
+    pipeline = DiffusionPipeline.from_pretrained(ckpt_id, vae=vae, torch_dtype=torch.bfloat16)
+    pipeline.transformer.to(memory_format=torch.channels_last)
+    pipeline.text_encoder.to(memory_format=torch.channels_last)
+    pipeline.text_encoder_2.to(memory_format=torch.channels_last)
+    pipeline.vae.to(memory_format=torch.channels_last)
+    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune")
+    pipeline.vae.enable_tiling()
+    pipeline._exclude_from_cpu_offload = ["vae"]
+    pipeline.enable_sequential_cpu_offload()
+
+    for _ in range(2):
+        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)
+    fart()
+    return pipeline
+    fart()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    fart()
+    generator = Generator("cuda").manual_seed(request.seed)
+    image=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]
+    return(image)