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 +1,34 @@
+[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",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "jokerbit/flux.1-schnell-Robert-int8wo"
+revision = "5ef0012f11a863e5111ec56540302a023bc8587b" 
+
+
+[[tool.edge-maxxing.models]]
+repository = "madebyollin/taef1"
+revision = "2d552378e58c9c94201075708d7de4e1163b2689"
+
+
+[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"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, 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 +1,84 @@
+import gc
+import os
+from typing import TypeAlias
+
+import torch
+from PIL.Image import Image
+from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderKL, AutoencoderTiny
+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
+
+
+Pipeline: TypeAlias = FluxPipeline
+torch.backends.cudnn.benchmark = True
+
+CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
+REVISION = "5ef0012f11a863e5111ec56540302a023bc8587b"
+
+TinyVAE = "madebyollin/taef1"
+TinyVAE_REV = "2d552378e58c9c94201075708d7de4e1163b2689"
+
+
+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)
+    vae = AutoencoderTiny.from_pretrained(
+            TinyVAE, 
+            revision=TinyVAE_REV, 
+            local_files_only=True,
+            torch_dtype=torch.bfloat16)
+    pipeline = FluxPipeline.from_pretrained(
+        CHECKPOINT,
+        revision=REVISION,
+        transformer=transformer,
+        vae=vae,
+        local_files_only=True,
+        torch_dtype=torch.bfloat16,
+    ).to("cuda")  
+
+#    pipeline.vae.to(memory_format=torch.channels_last) 
+    quantize_(pipeline.vae, int8_weight_only())
+#    pipeline.vae = torch.compile(pipeline.vae, mode="reduce-overhead")
+#    pipeline.to("cuda")
+    for _ in range(2):
+        pipeline("cat", num_inference_steps=4)
+
+    return pipeline
+
+@torch.inference_mode()
+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]
+
+
+if __name__ == "__main__":
+    from time import perf_counter
+    PROMPT = 'martyr, semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'
+    request = TextToImageRequest(prompt=PROMPT, 
+            height=None,
+            width=None,
+            seed=666)
+    start_time = perf_counter()
+    pipe_ = load_pipeline()
+    stop_time = perf_counter()
+    print(f"Pipeline is loaded in {stop_time - start_time}s")
+    for _ in range(4):
+        start_time = perf_counter()
+        infer(request, pipe_)
+        stop_time = perf_counter()
+        print(f"Request in {stop_time - start_time}s")