Upload folder using huggingface_hub

pyproject.toml

@@ -0,0 +1,29 @@
+[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",
+  "hf_transfer==0.1.8",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+  "setuptools>=75.3.0",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "jokerbit/flux.1-schnell-Robert-int8wo"
+revision = "5ef0012f11a863e5111ec56540302a023bc8587b" 
+
+[project.scripts]
+start_inference = "main:main"

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
+import torch
+from pipeline import load_pipeline, infer
+
+SOCKET = abspath(Path(__file__).parent.parent / "inferences.sock")
+
+
+def main():
+    print(f"Loading pipeline")
+    pipeline = load_pipeline()
+    generator = torch.Generator(pipeline.device)
+    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"accepted")
+
+            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 +1,87 @@
+import gc
+import os
+from typing import TypeAlias
+
+import torch
+from PIL.Image import Image
+from diffusers import (
+    FluxPipeline,
+    FluxTransformer2DModel,
+    AutoencoderKL,
+    AutoencoderTiny,
+    DiffusionPipeline,
+)
+from huggingface_hub.constants import HF_HUB_CACHE
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from torchao.quantization import quantize_, int8_weight_only
+from transformers import T5EncoderModel, CLIPTextModel, logging
+import torch._dynamo
+
+torch._dynamo.config.suppress_errors = True
+
+Pipeline: TypeAlias = FluxPipeline
+
+torch.backends.cudnn.benchmark = True
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+
+
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
+REVISION = "5ef0012f11a863e5111ec56540302a023bc8587b"
+
+
+def load_pipeline() -> Pipeline:
+    path = os.path.join(
+        HF_HUB_CACHE,
+        "models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b/transformer",
+    )
+    transformer = FluxTransformer2DModel.from_pretrained(
+        path, use_safetensors=False, local_files_only=True, torch_dtype=torch.bfloat16
+    )
+
+    pipeline = FluxPipeline.from_pretrained(
+        CHECKPOINT,
+        revision=REVISION,
+        transformer=transformer,
+        local_files_only=True,
+        torch_dtype=torch.bfloat16,
+    ).to("cuda")
+
+    pipeline.to(memory_format=torch.channels_last)
+    pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune")
+    quantize_(pipeline.vae, int8_weight_only())
+    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune")
+
+    with torch.no_grad():
+        for _ in range(5):
+            pipeline(
+                prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness",
+                width=1024,
+                height=1024,
+                guidance_scale=0.0,
+                num_inference_steps=4,
+                max_sequence_length=256,
+            )
+    torch.cuda.empty_cache()
+    return pipeline
+
+
+@torch.no_grad()
+def infer(
+    request: TextToImageRequest, pipeline: Pipeline, generator: torch.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,
+    ).images[0]