Initial commit

.gitattributes

@@ -0,0 +1,37 @@
+*.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
+RobertML.png filter=lfs diff=lfs merge=lfs -text
+backup.png filter=lfs diff=lfs merge=lfs -text

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", "RobertML/FLUX.1-schnell-int8wo", "city96/t5-v1_1-xxl-encoder-bf16","RobertML/FLUX.1-schnell-vae_int8"]
+
+[project.scripts]
+start_inference = "main:main"

src/flux_schnell_edge_inference.egg-info/PKG-INFO

@@ -0,0 +1,15 @@
+Metadata-Version: 2.1
+Name: flux-schnell-edge-inference
+Version: 7
+Summary: An edge-maxxing model submission for the 4090 Flux contest
+Requires-Python: <3.13,>=3.10
+Requires-Dist: diffusers==0.31.0
+Requires-Dist: transformers==4.46.2
+Requires-Dist: accelerate==1.1.0
+Requires-Dist: omegaconf==2.3.0
+Requires-Dist: torch==2.5.1
+Requires-Dist: protobuf==5.28.3
+Requires-Dist: sentencepiece==0.2.0
+Requires-Dist: edge-maxxing-pipelines@ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines
+Requires-Dist: gitpython>=3.1.43
+Requires-Dist: torchao>=0.6.1

src/flux_schnell_edge_inference.egg-info/SOURCES.txt

@@ -0,0 +1,10 @@
+README.md
+pyproject.toml
+src/main.py
+src/pipeline.py
+src/flux_schnell_edge_inference.egg-info/PKG-INFO
+src/flux_schnell_edge_inference.egg-info/SOURCES.txt
+src/flux_schnell_edge_inference.egg-info/dependency_links.txt
+src/flux_schnell_edge_inference.egg-info/entry_points.txt
+src/flux_schnell_edge_inference.egg-info/requires.txt
+src/flux_schnell_edge_inference.egg-info/top_level.txt

src/flux_schnell_edge_inference.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

src/flux_schnell_edge_inference.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+start_inference = main:main

src/flux_schnell_edge_inference.egg-info/requires.txt

@@ -0,0 +1,10 @@
+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

src/flux_schnell_edge_inference.egg-info/top_level.txt

@@ -0,0 +1,2 @@
+main
+pipeline

src/main.py

@@ -0,0 +1,58 @@
+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
+from git import Repo
+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 +1,126 @@
+import os
+import gc
+import torch
+from torch import Generator
+from PIL.Image import Image
+from diffusers import AutoencoderKL, FluxPipeline
+from diffusers.image_processor import VaeImageProcessor
+from pipelines.models import TextToImageRequest
+from transformers import T5EncoderModel
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:False,garbage_collection_threshold:0.001"
+torch.set_float32_matmul_precision("medium")
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+dtype = torch.bfloat16
+Pipeline = None
+# Configure CUDA settings
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.cuda.set_per_process_memory_fraction(0.999)
+
+class BasicQuantization:
+    def __init__(self, bits=1):
+        self.bits = bits
+        self.qmin = -(2**(bits-1))
+        self.qmax = 2**(bits-1) - 1
+
+    def quantize_tensor(self, tensor):
+        scale = (tensor.max() - tensor.min()) / (self.qmax - self.qmin)
+        zero_point = self.qmin - torch.round(tensor.min() / scale)
+        qtensor = torch.round(tensor / scale + zero_point)
+        qtensor = torch.clamp(qtensor, self.qmin, self.qmax)
+        return (qtensor - zero_point) * scale, scale, zero_point
+
+class ModelQuantization:
+    def __init__(self, model, bits=9):
+        self.model = model
+        self.quant = BasicQuantization(bits)
+
+    def quantize_model(self):
+        for name, module in self.model.named_modules():
+            if isinstance(module, torch.nn.Linear):
+                if hasattr(module, 'weightML'):
+                    quantized_weight, _, _ = self.quant.quantize_tensor(module.weight)
+                    module.weight = torch.nn.Parameter(quantized_weight)
+                if hasattr(module, 'bias') and module.bias is not None:
+                    quantized_bias, _, _ = self.quant.quantize_tensor(module.bias)
+                    module.bias = torch.nn.Parameter(quantized_bias)
+
+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:
+    empty_cache()
+    
+    # Load and quantize VAE
+    vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=dtype)
+    quantizer = ModelQuantization(vae)
+    quantizer.quantize_model()
+    
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=torch.bfloat16
+    )
+
+    # Initialize pipeline
+    pipeline = FluxPipeline.from_pretrained(
+        ckpt_id,
+        text_encoder_2=text_encoder_2,
+        vae=vae,
+        torch_dtype=dtype
+    )
+
+
+    # Optimize memory format
+    for component in [pipeline.text_encoder, pipeline.text_encoder_2, pipeline.transformer, pipeline.vae]:
+        component.to(memory_format=torch.channels_last)
+
+    # Compile and configure pipeline
+    pipeline.vae = torch.compile(pipeline.vae, fullgraph=True, dynamic=False, mode="max-autotune")
+    pipeline._exclude_from_cpu_offload = ["vae"]
+    pipeline.enable_sequential_cpu_offload()
+
+    # Warmup run
+    empty_cache()
+    for _ in range(3):
+        pipeline(
+            prompt="posteroexternal, eurythmical, inspection, semicotton, specification, Mercatorial, ethylate, misprint",
+            width=1024,
+            height=1024,
+            guidance_scale=0.0,
+            num_inference_steps=4,
+            max_sequence_length=256
+        )
+    
+    return pipeline
+
+_inference_count = 0
+
+@torch.inference_mode()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    global _inference_count
+    
+    # Clear on first inference
+    if _inference_count == 0:
+        empty_cache()
+    
+    # Increment counter and empty cache every 4 inferences
+    _inference_count += 1
+    if _inference_count >= 4:
+        empty_cache()
+        _inference_count = 0
+    
+    torch.cuda.reset_peak_memory_stats()
+    generator = Generator("cuda").manual_seed(request.seed)
+    return pipeline(
+            prompt=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]