Initial commit

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+backup.png filter=lfs diff=lfs merge=lfs -text

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`

decoder.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7e85a093942d8d5fb785867065bd9dd1a38eb51e46af7bd5e24e0ef530b5bd9
+size 2494680

encoder.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad2516494613cedb1af1bff489f597b030c072bd8cc47fed8f2fe4e8575948bd
+size 2494552

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",
+  "para_attn"
+]
+
+[[tool.edge-maxxing.models]]
+repository = "manbeast3b/Flux.1.schnell-quant2"
+revision = "44eb293715147878512da10bf3bc47cd14ec8c55" 
+
+[[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 +1,57 @@
+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, float8_weight_only
+from transformers import T5EncoderModel, CLIPTextModel, logging 
+from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe as caching
+import gc
+import os
+from typing import TypeAlias
+import torch
+
+Pipeline = False
+
+torch._inductor.config.epilogue_fusion = False
+torch._dynamo.config.suppress_errors = True
+
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.coordinate_descent_check_all_directions = True
+torch._inductor.config.conv_1x1_as_mm = True
+
+torch.backends.cudnn.benchmark = True
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+
+BASE_CHECKPOINT = "manbeast3b/Flux.1.schnell-quant2"
+BASE_REVISION = "44eb293715147878512da10bf3bc47cd14ec8c55"
+
+def load_pipeline() -> Pipeline:
+    
+    path = os.path.join(HF_HUB_CACHE, "models--manbeast3b--Flux.1.schnell-quant2/snapshots/44eb293715147878512da10bf3bc47cd14ec8c55/transformer")
+    mod_transformer = FluxTransformer2DModel.from_pretrained(path, use_safetensors=False,local_files_only=True,torch_dtype=torch.bfloat16)
+    
+    vae = AutoencoderTiny.from_pretrained(
+            "madebyollin/taef1", 
+            revision="2d552378e58c9c94201075708d7de4e1163b2689", 
+            local_files_only=True,
+            torch_dtype=torch.bfloat16)
+    vae.encoder.load_state_dict(torch.load("encoder.pth"), strict=False); vae.decoder.load_state_dict(torch.load("decoder.pth"), strict=False)
+    
+    pipeline = FluxPipeline.from_pretrained(BASE_CHECKPOINT,revision=BASE_REVISION,transformer=mod_transformer,vae=vae,local_files_only=True,torch_dtype=torch.bfloat16).to("cuda")
+    pipeline.to(memory_format=torch.channels_last)
+    # quantize_(pipeline.vae, float8_weight_only()) 
+    
+    quantize_(pipeline.vae, int8_weight_only()) 
+    pipeline = caching(pipeline,residual_diff_threshold=0.86)
+    
+    with torch.inference_mode():
+        for _ in range(2):
+            pipeline("Eagles, Chiefs, Kendrick Lamar, Halftime Show, DunKings, Clydesdales, Tush Push, Moneyline, Taylor Swift, Serena Williams", num_inference_steps=4, guidance_scale=0.0,)
+    
+    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]