Upload folder using huggingface_hub

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",
+  "setuptools>=75.3.0",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+
+[[tool.edge-maxxing.models]]
+repository = "city96/t5-v1_1-xxl-encoder-bf16"
+revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+
+[[tool.edge-maxxing.models]]
+repository = "park234/Flux1-schnell-int8"
+revision = "ba40332fa9ca3e25d439616fca8ac934a96b73f7"
+
+[[tool.edge-maxxing.models]]
+repository = "park234/flux1-schnell-vae32"
+revision = "d7c57ba8ee0d581d67256219c152a8e69da853b5"
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/main.py

@@ -0,0 +1,57 @@
+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 +1,85 @@
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+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._dynamo.config.suppress_errors = True
+
+Pipeline = None
+
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+
+
+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()
+
+    transformer_path = os.path.join(
+        HF_HUB_CACHE,
+        "models--park234--Flux1-schnell-int8/snapshots/ba40332fa9ca3e25d439616fca8ac934a96b73f7",
+    )
+    transformer_model = FluxTransformer2DModel.from_pretrained(
+        transformer_path, torch_dtype=torch.bfloat16, use_safetensors=False
+    ).to(memory_format=torch.channels_last)
+    
+    vae_model = AutoencoderTiny.from_pretrained(
+        "park234/flux1-schnell-vae32",
+        revision="d7c57ba8ee0d581d67256219c152a8e69da853b5",
+        torch_dtype=torch.bfloat16,
+    )
+    
+    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)
+    
+    pipeline = FluxPipeline.from_pretrained(
+        ckpt_id,
+        vae=vae_model,
+        revision=ckpt_revision,
+        transformer=transformer_model,
+        text_encoder_2=text_encoder_2,
+        torch_dtype=torch.bfloat16,
+    ).to("cuda")
+    
+    pipeline.transformer = torch.compile(pipeline.transformer, mode="reduce-overhead")
+    
+    empty_cache()
+    return pipeline
+
+
+@torch.no_grad()
+def infer(
+    request: TextToImageRequest, pipeline: Pipeline, generator: Generator
+) -> Image:
+    return 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]