Initial commit

.gitattributes

@@ -0,0 +1 @@
+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`

base-flux-simpleflux9

@@ -0,0 +1 @@
+Subproject commit fa96c90df9ee60abf4d096a5e5659c8d15ee49dd

pyproject.toml

@@ -0,0 +1,27 @@
+[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",
+  "gitpython>=3.1.43",
+  "torchao>=0.6.1"
+]
+
+[tool.edge-maxxing]
+models = ["black-forest-labs/FLUX.1-schnell", "madebyollin/taef1"]
+
+[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 +1,70 @@
+from diffusers import AutoencoderKL, AutoencoderTiny
+from diffusers.image_processor import VaeImageProcessor
+import torch
+import torch._dynamo
+import gc
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from diffusers import FluxPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+
+Pipeline = None
+MODEL_ID = "black-forest-labs/FLUX.1-schnell"
+DTYPE = torch.bfloat16
+def clear():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def load_pipeline() -> Pipeline:    
+    clear()
+    # vae = AutoencoderKL.from_pretrained(
+    #     MODEL_ID, subfolder="vae", torch_dtype=torch.bfloat16
+    # )
+    vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=DTYPE)
+    # quantize_(vae, fpx_weight_only(3, 2))
+    quantize_(vae, int8_weight_only())
+    pipeline = FluxPipeline.from_pretrained(MODEL_ID,vae=vae,
+                                        torch_dtype=DTYPE)
+    torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.cuda.set_per_process_memory_fraction(0.99)
+    pipeline.text_encoder.to(memory_format=torch.channels_last)
+    pipeline.text_encoder_2.to(memory_format=torch.channels_last)
+    pipeline.transformer.to(memory_format=torch.channels_last)
+    pipeline.vae.to(memory_format=torch.channels_last)
+    pipeline.vae = torch.compile(pipeline.vae)
+    pipeline._exclude_from_cpu_offload = ["vae"]
+    pipeline.enable_sequential_cpu_offload()
+    clear()
+    for _ in range(1):
+        pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.1, num_inference_steps=4, max_sequence_length=256)
+    return pipeline
+
+# sample = True
+# @torch.inference_mode()
+# def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+#     global sample
+#     if sample:
+#         clear()
+#         sample = None
+#     # torch.cuda.reset_peak_memory_stats()
+#     generator = Generator("cuda").manual_seed(request.seed)
+#     image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]
+#     return(image)
+
+sample = True
+@torch.inference_mode()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    global sample
+    if sample:
+        clear()
+        sample = None
+    # torch.cuda.reset_peak_memory_stats()
+    generator = Generator("cuda").manual_seed(request.seed)
+    image = None
+    with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+        image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]
+    return(image)