Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* 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
+sample.png filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,9 @@
+**/.cache
+**/__pycache__
+**/*.egg-info
+*.safetensors
+**/.venv
+.venv
+.git
+*.png
+*.jpeg

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 +1,28 @@
+[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",
+  "torchao>=0.6.1",
+  "ipython>=8.29.0",
+  "setuptools >= 75.0"
+]
+
+[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 +1,134 @@
+import os
+from diffusers import FluxPipeline, AutoencoderKL, FluxTransformer2DModel
+from diffusers.image_processor import VaeImageProcessor
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel, CLIPTextConfig, T5Config
+import torch
+import gc
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from torchao.quantization import quantize_, int8_weight_only, int8_dynamic_activation_int8_weight
+from time import perf_counter
+
+
+HOME = os.environ["HOME"]
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+FLUX_CHECKPOINT = "black-forest-labs/FLUX.1-schnell"
+# QUANTIZED_MODEL = []
+QUANTIZED_MODEL = ["transformer", "text_encoder_2", "text_encoder", "vae"]
+
+
+QUANT_CONFIG = int8_weight_only()
+DTYPE = torch.bfloat16
+NUM_STEPS = 4
+
+def get_transformer(quantize: bool = True, quant_config = int8_weight_only(), quant_ckpt: str = None):
+    if quant_ckpt is not None:
+        config = FluxTransformer2DModel.load_config(FLUX_CHECKPOINT, subfolder="transformer")
+        model = FluxTransformer2DModel.from_config(config).to(DTYPE)
+        state_dict = torch.load(quant_ckpt, map_location="cpu")
+        model.load_state_dict(state_dict, assign=True)
+        print(f"Loaded {quant_ckpt}")
+        return model
+    
+    model = FluxTransformer2DModel.from_pretrained(
+        FLUX_CHECKPOINT, subfolder="transformer", torch_dtype=DTYPE
+        )
+    if quantize:
+        quantize_(model, quant_config)
+    return model
+
+
+def get_text_encoder(quantize: bool = True, quant_config = int8_weight_only(), quant_ckpt: str = None):
+    if quant_ckpt is not None:
+        config = CLIPTextConfig.from_pretrained(FLUX_CHECKPOINT, subfolder="text_encoder")
+        model = CLIPTextModel(config).to(DTYPE)
+        state_dict = torch.load(quant_ckpt, map_location="cpu")
+        model.load_state_dict(state_dict, assign=True)
+        print(f"Loaded {quant_ckpt}")
+        return model
+    
+    model = CLIPTextModel.from_pretrained(
+        FLUX_CHECKPOINT, subfolder="text_encoder", torch_dtype=DTYPE
+        )
+    if quantize:
+        quantize_(model, quant_config)
+    return model
+
+
+def get_text_encoder_2(quantize: bool = True, quant_config = int8_weight_only(), quant_ckpt: str = None):
+    if quant_ckpt is not None:
+        config = T5Config.from_pretrained(FLUX_CHECKPOINT, subfolder="text_encoder_2")
+        model = T5EncoderModel(config).to(DTYPE)
+        state_dict = torch.load(quant_ckpt, map_location="cpu")
+        print(f"Loaded {quant_ckpt}")
+        model.load_state_dict(state_dict, assign=True)
+        return model
+    
+    model = T5EncoderModel.from_pretrained(
+        FLUX_CHECKPOINT, subfolder="text_encoder_2", torch_dtype=DTYPE
+        )
+    if quantize:
+        quantize_(model, quant_config)
+    return model
+
+
+def get_vae(quantize: bool = True, quant_config = int8_weight_only(), quant_ckpt: str = None):
+    if quant_ckpt is not None:
+        config = AutoencoderKL.load_config(FLUX_CHECKPOINT, subfolder="vae")
+        model = AutoencoderKL.from_config(config).to(DTYPE)
+        state_dict = torch.load(quant_ckpt, map_location="cpu")
+        model.load_state_dict(state_dict, assign=True)
+        print(f"Loaded {quant_ckpt}")
+        return model
+    model = AutoencoderKL.from_pretrained(
+        FLUX_CHECKPOINT, subfolder="vae", torch_dtype=DTYPE
+        )
+    if quantize:
+        quantize_(model, quant_config)
+    return model
+
+
+def empty_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+
+def load_pipeline() -> FluxPipeline:
+    empty_cache()
+    
+    pipe = FluxPipeline.from_pretrained(FLUX_CHECKPOINT,
+                                        torch_dtype=DTYPE)
+    
+    pipe.text_encoder_2.to(memory_format=torch.channels_last)
+    pipe.transformer.to(memory_format=torch.channels_last)
+    pipe.vae.to(memory_format=torch.channels_last)
+    pipe.vae = torch.compile(pipe.vae)
+    
+    pipe._exclude_from_cpu_offload = ["vae"]
+    
+    pipe.enable_sequential_cpu_offload()
+    
+    empty_cache()
+    pipe("cat", guidance_scale=0., max_sequence_length=256, num_inference_steps=4)
+    return pipe
+
+@torch.inference_mode()
+def infer(request: TextToImageRequest, _pipeline: FluxPipeline) -> Image:
+    if request.seed is None:
+        generator = None
+    else:
+        generator = Generator(device="cuda").manual_seed(request.seed)
+
+    empty_cache()
+    image = _pipeline(prompt=request.prompt,
+                      width=request.width,
+                      height=request.height,
+                      guidance_scale=0.0,
+                      generator=generator,
+                      output_type="pil",
+                      max_sequence_length=256,
+                      num_inference_steps=NUM_STEPS).images[0]
+    return image