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 +1,25 @@
+[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 = "7"
+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",
+]
+
+[tool.edge-maxxing]
+models = ["black-forest-labs/FLUX.1-schnell"]
+
+[project.scripts]
+start_inference = "main:main"

src/main.py

@@ -0,0 +1,59 @@
+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")
+
+            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 +1,116 @@
+from diffusers import FluxPipeline, AutoencoderKL
+from diffusers.image_processor import VaeImageProcessor
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import gc
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+
+Pipeline = None
+
+CHECKPOINT = "black-forest-labs/FLUX.1-schnell"
+
+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:
+    infer(TextToImageRequest(prompt=""), Pipeline)
+
+    return Pipeline
+
+
+def encode_prompt(prompt: str):
+    text_encoder = CLIPTextModel.from_pretrained(
+        CHECKPOINT,
+        subfolder="text_encoder",
+        torch_dtype=torch.bfloat16,
+    )
+
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        CHECKPOINT,
+        subfolder="text_encoder_2",
+        torch_dtype=torch.bfloat16,
+    )
+
+    tokenizer = CLIPTokenizer.from_pretrained(CHECKPOINT, subfolder="tokenizer")
+    tokenizer_2 = T5TokenizerFast.from_pretrained(CHECKPOINT, subfolder="tokenizer_2")
+
+    pipeline = FluxPipeline.from_pretrained(
+        CHECKPOINT,
+        text_encoder=text_encoder,
+        text_encoder_2=text_encoder_2,
+        tokenizer=tokenizer,
+        tokenizer_2=tokenizer_2,
+        transformer=None,
+        vae=None,
+    ).to("cuda")
+
+    with torch.no_grad():
+        return pipeline.encode_prompt(
+            prompt=prompt,
+            prompt_2=None,
+            max_sequence_length=256,
+        )
+
+
+def infer_latents(prompt_embeds, pooled_prompt_embeds, width: int | None, height: int | None, seed: int | None):
+    pipeline = FluxPipeline.from_pretrained(
+        CHECKPOINT,
+        text_encoder=None,
+        text_encoder_2=None,
+        tokenizer=None,
+        tokenizer_2=None,
+        vae=None,
+        torch_dtype=torch.bfloat16,
+    ).to("cuda")
+
+    if seed is None:
+        generator = None
+    else:
+        generator = Generator(pipeline.device).manual_seed(seed)
+
+    return pipeline(
+        prompt_embeds=prompt_embeds,
+        pooled_prompt_embeds=pooled_prompt_embeds,
+        num_inference_steps=4,
+        guidance_scale=0.0,
+        width=width,
+        height=height,
+        generator=generator,
+        output_type="latent",
+    ).images
+
+
+def infer(request: TextToImageRequest, _pipeline: Pipeline) -> Image:
+    empty_cache()
+
+    prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(request.prompt)
+
+    empty_cache()
+
+    latents = infer_latents(prompt_embeds, pooled_prompt_embeds, request.width, request.height, request.seed)
+
+    empty_cache()
+
+    vae = AutoencoderKL.from_pretrained(
+        CHECKPOINT,
+        subfolder="vae",
+        torch_dtype=torch.bfloat16,
+    ).to("cuda")
+
+    vae_scale_factor = 2 ** (len(vae.config.block_out_channels))
+    image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor)
+
+    height = request.height or 64 * vae_scale_factor
+    width = request.width or 64 * vae_scale_factor
+
+    with torch.no_grad():
+        latents = FluxPipeline._unpack_latents(latents, height, width, vae_scale_factor)
+        latents = (latents / vae.config.scaling_factor) + vae.config.shift_factor
+
+        image = vae.decode(latents, return_dict=False)[0]
+        return image_processor.postprocess(image, output_type="pil")[0]