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 +1,28 @@
+[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",
+  "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",
+  "setuptools >= 75.0"
+]
+
+[tool.edge-maxxing]
+models = ["black-forest-labs/FLUX.1-schnell"]
+
+[project.scripts]
+start_inference = "main:main"

src/main.py

@@ -0,0 +1,62 @@
+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 +1,81 @@
+import os
+from diffusers import FluxPipeline, AutoencoderKL, FluxTransformer2DModel
+from diffusers.image_processor import VaeImageProcessor
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel, CLIPTextConfig, T5Config
+import torch
+import gc
+from PIL import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from time import perf_counter
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+
+class EightQuantize:
+    def __init__(self, bits=8):
+        self.bits = bits
+        self.qmax = (1 << bits) - 1
+        
+    def __call__(self, x):
+        scale = x.max() / self.qmax
+        x_quant = torch.clip(torch.round(x / scale), 0, self.qmax)
+        return x_quant * scale
+
+
+CHECKPOINT = "black-forest-labs/FLUX.1-schnell"
+DTYPE = torch.bfloat16
+NUM_STEPS = 4
+
+def empty_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def load_pipeline() -> FluxPipeline:
+    empty_cache()
+    is_quantize = 0
+    _pipe = None
+    pipe = FluxPipeline.from_pretrained(CHECKPOINT, torch_dtype=DTYPE)
+    
+    pipe.text_encoder_2.to(memory_format=torch.channels_last)
+    pipe.transformer.to(memory_format=torch.channels_last)
+    
+    pipe.vae.to(memory_format=torch.channels_last)
+    pipe.vae = torch.compile(pipe.vae)
+    pipe._exclude_from_cpu_offload = ["vae"]
+    
+    try:
+        if is_quantize:
+            quantizer = EightQuantize()
+            with torch.no_grad():
+                for param in _pipe.vae.parameters():
+                    param.data = quantizer(param.data)
+    except Exception as e:
+        print(f"Quantization warning: {e}")
+    
+    pipe.enable_sequential_cpu_offload()
+    
+    empty_cache()
+    pipe("dog", guidance_scale=0.0, max_sequence_length=256, num_inference_steps=4)
+    return pipe
+
+@torch.inference_mode()
+def infer(request: TextToImageRequest, _pipeline: FluxPipeline) -> Image:
+    torch.cuda.reset_peak_memory_stats()
+    
+    if request.seed is None:
+        generator = None
+    else:
+        generator = Generator(device="cuda").manual_seed(request.seed)
+
+    empty_cache()
+    image = _pipeline(prompt=request.prompt,
+                      width=request.width,
+                      height=request.height,
+                      guidance_scale=0.0,
+                      generator=generator,
+                      output_type="pil",
+                      max_sequence_length=256,
+                      num_inference_steps=NUM_STEPS).images[0]
+    return image