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 +1,27 @@
+[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",
+  "gitpython>=3.1.43",
+  "torchao>=0.6.1",
+]
+
+[tool.edge-maxxing]
+models = ["black-forest-labs/FLUX.1-schnell", "city96/t5-v1_1-xxl-encoder-bf16", "madebyollin/taef1"]
+
+[project.scripts]
+start_inference = "main:main"

src/flux_schnell_edge_inference.egg-info/PKG-INFO

@@ -0,0 +1,15 @@
+Metadata-Version: 2.1
+Name: flux-schnell-edge-inference
+Version: 7
+Summary: An edge-maxxing model submission for the 4090 Flux contest
+Requires-Python: <3.13,>=3.10
+Requires-Dist: diffusers==0.31.0
+Requires-Dist: transformers==4.46.2
+Requires-Dist: accelerate==1.1.0
+Requires-Dist: omegaconf==2.3.0
+Requires-Dist: torch==2.5.1
+Requires-Dist: protobuf==5.28.3
+Requires-Dist: sentencepiece==0.2.0
+Requires-Dist: edge-maxxing-pipelines@ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines
+Requires-Dist: gitpython>=3.1.43
+Requires-Dist: torchao>=0.6.1

src/flux_schnell_edge_inference.egg-info/SOURCES.txt

@@ -0,0 +1,10 @@
+README.md
+pyproject.toml
+src/main.py
+src/pipeline.py
+src/flux_schnell_edge_inference.egg-info/PKG-INFO
+src/flux_schnell_edge_inference.egg-info/SOURCES.txt
+src/flux_schnell_edge_inference.egg-info/dependency_links.txt
+src/flux_schnell_edge_inference.egg-info/entry_points.txt
+src/flux_schnell_edge_inference.egg-info/requires.txt
+src/flux_schnell_edge_inference.egg-info/top_level.txt

src/flux_schnell_edge_inference.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

src/flux_schnell_edge_inference.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+start_inference = main:main

src/flux_schnell_edge_inference.egg-info/requires.txt

@@ -0,0 +1,10 @@
+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
+torchao>=0.6.1

src/flux_schnell_edge_inference.egg-info/top_level.txt

@@ -0,0 +1,2 @@
+main
+pipeline

src/main.py

@@ -0,0 +1,58 @@
+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
+from git import Repo
+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 @@
+from diffusers import AutoencoderKL
+
+from transformers import T5EncoderModel
+import torch
+import torch._dynamo
+import gc
+from PIL import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+import os
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False,garbage_collection_threshold:0.01"
+
+Pipeline = None
+MODEL_ID = "black-forest-labs/FLUX.1-schnell"
+DTYPE = torch.bfloat16
+def clear():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def load_pipeline() -> Pipeline:    
+    clear()
+   
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=DTYPE
+    )
+    vae=AutoencoderKL.from_pretrained(MODEL_ID, subfolder="vae", torch_dtype=DTYPE)
+    pipeline = DiffusionPipeline.from_pretrained(
+        MODEL_ID,
+        vae=vae,
+        text_encoder_2 = text_encoder_2,
+        torch_dtype=DTYPE,
+        )
+    
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cudnn.deterministic = False
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.cuda.set_per_process_memory_fraction(0.98)
+    pipeline.text_encoder.to(memory_format=torch.channels_last)
+    pipeline.transformer.to(memory_format=torch.channels_last)
+    pipeline.vae.to(memory_format=torch.channels_last)
+    pipeline.vae = torch.compile(pipeline.vae)
+    
+    pipeline._exclude_from_cpu_offload = ["vae"]
+    pipeline.enable_sequential_cpu_offload()
+    for _ in range(2):
+        pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    
+    return pipeline
+
+
+sample = True
+@torch.inference_mode()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    global sample
+    if sample:
+        clear()
+        sample = None
+    torch.cuda.reset_peak_memory_stats()
+    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)