Upload folder using huggingface_hub

README.md

@@ -0,0 +1,19 @@
+# flux-schnell-edge-inference
+
+This holds the baseline for the FLUX Schnel NVIDIA GeForce RTX 4090 contest, which can be forked freely and optimized
+
+Some recommendations are as follows:
+- Installing dependencies should be done in `pyproject.toml`, including git dependencies
+- HuggingFace models should be specified in the `models` array in the `pyproject.toml` file, and will be downloaded before benchmarking
+    - The pipeline does **not** have internet access so all dependencies and models must be included in the `pyproject.toml`
+    - Compiled models should be hosted on HuggingFace and included in the `models` array in the `pyproject.toml` (rather than compiling during loading). Loading time matters far more than file sizes
+- Avoid changing `src/main.py`, as that includes mostly protocol logic. Most changes should be in `models` and `src/pipeline.py`
+- Ensure the entire repository (excluding dependencies and HuggingFace models) is under 16MB
+
+For testing, you need a docker container with pytorch and ubuntu 22.04.
+You can download your listed dependencies with `uv`, installed with:
+```bash
+pipx ensurepath
+pipx install uv
+```
+You can then relock with `uv lock`, and then run with `uv run start_inference`

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",
+  "setuptools==75.2.0",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "slobers/Flux.1.Schnella"
+revision = "e34d670e44cecbbc90e4962e7aada2ac5ce8b55b"
+
+[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
+
+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 +1,53 @@
+#5.1
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import torch._dynamo
+import gc
+import os
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+
+Pipeline = None
+ids = "slobers/Flux.1.Schnella"
+Revision = "e34d670e44cecbbc90e4962e7aada2ac5ce8b55b"
+
+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()
+    path = os.path.join(HF_HUB_CACHE, "models--slobers--Flux.1.Schnella/snapshots/e34d670e44cecbbc90e4962e7aada2ac5ce8b55b/transformer")
+    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False)
+    pipeline = FluxPipeline.from_pretrained(ids, revision=Revision, transformer=transformer, local_files_only=True, torch_dtype=torch.bfloat16,)
+    pipeline.to("cuda")
+    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune", fullgraph=True, dynamic=True)
+    for _ in range(3):
+        pipeline(prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus", width=1024, height=1024, guidance_scale=5.0, num_inference_steps=4, max_sequence_length=256)
+    return pipeline
+    empty_cache()
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    generator = Generator(pipeline.device).manual_seed(request.seed)
+
+    return pipeline(
+        request.prompt,
+        generator=generator,
+        guidance_scale=6.5,
+        num_inference_steps=4,
+        max_sequence_length=256,
+        height=request.height,
+        width=request.width,
+    ).images[0]