Initial commit with folder contents

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* 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
+RobertML.png filter=lfs diff=lfs merge=lfs -text

pyproject.toml

@@ -0,0 +1,44 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "An edge-maxxing model submission by RobertML 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",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer"]
+
+[[tool.edge-maxxing.models]]
+repository = "RobertML/FLUX.1-schnell-int8wo"
+revision = "307e0777d92df966a3c0f99f31a6ee8957a9857a"
+
+[[tool.edge-maxxing.models]]
+repository = "city96/t5-v1_1-xxl-encoder-bf16"
+revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+
+[[tool.edge-maxxing.models]]
+repository = "proact/PRO_FLUX_9"
+revision = "0c58d671a73b6ed093b5064d453c4cf2b3f13183"
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/main.py

@@ -0,0 +1,54 @@
+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")
+            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 +1,73 @@
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer
+import torch
+import torch._dynamo
+import gc
+from PIL import Image as img
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+import time
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+import os
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+
+
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.enabled = True
+torch.backends.cudnn.benchmark = True
+
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9" 
+
+
+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:    
+    empty_cache()
+
+    dtype, device = torch.bfloat16, "cuda"
+    
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16", revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86", torch_dtype=torch.bfloat16
+    ).to(memory_format=torch.channels_last)
+    
+
+    path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
+    model = FluxTransformer2DModel.from_pretrained(path, torch_dtype=dtype, use_safetensors=False).to(memory_format=torch.channels_last)
+    pipeline = DiffusionPipeline.from_pretrained(
+        ckpt_id,
+        revision=ckpt_revision,
+        transformer=model,
+        text_encoder_2=text_encoder_2,
+        torch_dtype=dtype,
+        ).to(device)
+    quantize_(pipeline.vae, int8_weight_only())
+
+    pipeline(prompt="imprisonable, forechamber, demagogic, monotropic, blandiloquious, blechnoid, uncarnivorous", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    
+    empty_cache()
+    return pipeline
+
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
+    
+    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)