Initial commit

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm 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
+

pyproject.toml

@@ -0,0 +1,43 @@
+[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",
+  "optimum-quanto",
+  "hf_transfer==0.1.8",
+  "setuptools==75.2.0",
+  "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 = "RichardWilliam/XULF_T5_bf16"
+revision = "63a3d9ef7b586655600ac9bd4e4747d038237761"
+
+[[tool.edge-maxxing.models]]
+repository = "RichardWilliam/XULF_Vae"
+revision = "3ee225c539465c27adadec45c6e8af50a7397b7d"
+
+[[tool.edge-maxxing.models]]
+repository = "RichardWilliam/XULF_Transfomer"
+revision = "6860c51af40329808f270e159a0d018559a1204f"
+
+[project.scripts]
+start_inference = "main:main"

quantization_map.json

@@ -0,0 +1 @@
+{"encoder.block.0.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.0.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.0.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.0.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.0.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.0.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.0.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.1.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.2.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.3.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.4.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.5.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.6.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.7.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.8.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.9.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.10.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.11.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.12.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.13.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.14.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.15.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.16.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.17.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.18.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.19.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.20.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.21.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.22.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.0.SelfAttention.q": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.0.SelfAttention.k": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.0.SelfAttention.v": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.0.SelfAttention.o": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.1.DenseReluDense.wi_0": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.1.DenseReluDense.wi_1": {"weights": "qint8", "activations": "none"}, "encoder.block.23.layer.1.DenseReluDense.wo": {"weights": "qint8", "activations": "none"}}

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 +1,103 @@
+# FLux Optimization Pipeline
+import os
+import torch
+import torch._dynamo
+import gc
+
+
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+
+from PIL.Image import Image
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from pipelines.models import TextToImageRequest
+from optimum.quanto import requantize
+import json
+import transformers
+
+
+
+torch._dynamo.config.suppress_errors = True
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+
+CHECKPOINT = "black-forest-labs/FLUX.1-schnell"
+REVISION = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+Pipeline = None
+apply_quanto=1
+
+def reset_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def load_quanto_text_encoder_2(text_repo_path):
+    with open("quantization_map.json", "r") as f:
+        quantization_map = json.load(f)
+    with open(os.path.join(text_repo_path, "config.json"), "r") as f:
+        t5_config = transformers.T5Config(**json.load(f))
+    with torch.device("meta"):
+        text_encoder_2 = transformers.T5EncoderModel(t5_config).to(torch.bfloat16)
+    state_dict = None
+    requantize(text_encoder_2, state_dict, quantization_map, device=torch.device("cuda"))
+    return text_encoder_2
+
+def load_pipeline() -> Pipeline:
+
+    try:
+        text_repo_path = os.path.join(HF_HUB_CACHE, "models--RichardWilliam--XULF_T5_bf16/snapshots/63a3d9ef7b586655600ac9bd4e4747d038237761")
+        text_encoder_2 = load_quanto_text_encoder_2(text_repo_path=text_repo_path)
+    except:
+        text_encoder_2 =  T5EncoderModel.from_pretrained("RichardWilliam/XULF_T5_bf16", 
+                    revision = "63a3d9ef7b586655600ac9bd4e4747d038237761", 
+                    torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+
+    origin_vae = AutoencoderTiny.from_pretrained("RichardWilliam/XULF_Vae",
+                    revision="3ee225c539465c27adadec45c6e8af50a7397b7d", 
+                    torch_dtype=torch.bfloat16)
+
+
+    main_path = os.path.join(HF_HUB_CACHE, "models--RichardWilliam--XULF_Transfomer/snapshots/6860c51af40329808f270e159a0d018559a1204f")
+    origin_trans = FluxTransformer2DModel.from_pretrained(main_path,
+                        torch_dtype=torch.bfloat16, 
+                        use_safetensors=False).to(memory_format=torch.channels_last)
+    transformer = origin_trans
+
+    pipeline = DiffusionPipeline.from_pretrained(CHECKPOINT, 
+                        revision=REVISION, 
+                        vae=origin_vae, 
+                        transformer=transformer, 
+                        text_encoder_2=text_encoder_2, 
+                        torch_dtype=torch.bfloat16)
+    pipeline.to("cuda")
+
+    for __ in range(3):
+        pipeline(prompt="sweet, subordinative, gender, mormyre, arteriolosclerosis, positivism, Antiochianism, palmerite", 
+                        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:
+
+    reset_cache()
+
+    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]