Initial commit with folder contents

pyproject.toml

@@ -0,0 +1,43 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "HitmanReborn"
+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",
+  "setuptools==75.2.0",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer", "vae", "text_encoder_2"]
+
+[[tool.edge-maxxing.models]]
+repository = "db900/axis-morph"
+revision = "f0981b786fdc1bf6b398ad06658ab0776ba047ec"
+
+[[tool.edge-maxxing.models]]
+repository = "db900/neural-lattice"
+revision = "31581dabff21433df68d22d5539d07de6a87380a"
+
+[[tool.edge-maxxing.models]]
+repository = "db900/trans-flux"
+revision = "2632cc4202aa3e7f459031cc45804e3693da6722"
+
+
+[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
+
+from pipeline import load_pipeline, infer
+
+SOCKET = abspath(Path(__file__).parent.parent / "inferences.sock")
+
+
+def main():
+    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 +1,87 @@
+from PIL.Image import Image
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel
+from PIL.Image import Image
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from PIL.Image import Image
+from diffusers import AutoencoderTiny
+from pipelines.models import TextToImageRequest
+import os
+import torch
+import torch._dynamo
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+
+Pipeline = None
+basePT = "forswearer, skullcap, Juglandales, bluelegs, cunila, carbro, Ammonites"
+
+class Quantization:
+    def __init__(self, model):
+        self.model = model
+        self.layer_configs = {
+            "single_transformer_blocks.0.attn.norm_k.weight": (128, 0.96),
+            "single_transformer_blocks.0.attn.norm_q.weight": (128, 0.96),
+            "single_transformer_blocks.0.attn.norm_v.weight": (128, 0.96)
+        }
+
+    def apply(self):
+        for name, param in self.model.named_parameters():
+            if param.requires_grad:
+                layer_name = name.split(".")[0]
+                if layer_name in self.layer_configs:
+                    num_bins, scale_factor = self.layer_configs[layer_name]
+                    with torch.no_grad():
+                        # Normalize weights, apply binning, and rescale
+                        param_min = param.min()
+                        param_max = param.max()
+                        param_range = param_max - param_min
+
+                        if param_range > 0:
+                            normalized = (param - param_min) / param_range
+                            binned = torch.round(normalized * (num_bins - 1)) / (num_bins - 1)
+                            rescaled = binned * param_range + param_mins
+                            params.data.copy_(rescaled * scale_factor)
+                        else:
+                            params.data.zero_()
+
+        return self.model
+
+def load_pipeline() -> Pipeline:
+
+    text_encoder_2 = T5EncoderModel.from_pretrained("db900/neural-lattice", revision = "31581dabff21433df68d22d5539d07de6a87380a", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+    vae = AutoencoderTiny.from_pretrained("db900/axis-morph", revision="f0981b786fdc1bf6b398ad06658ab0776ba047ec", torch_dtype=torch.bfloat16)
+    default = FluxTransformer2DModel.from_pretrained(os.path.join(HF_HUB_CACHE, "models--db900--trans-flux/snapshots/2632cc4202aa3e7f459031cc45804e3693da6722"), torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
+
+    try:
+        transformer = Quantization(transformer).apply()
+    except Exception as e:
+        transformer = default
+
+    pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", revision="741f7c3ce8b383c54771c7003378a50191e9efe9", vae=vae, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16)
+    pipeline.to("cuda")
+
+    for _ in range(3):
+        pipeline(prompt=basePT, width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    
+    return pipeline
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    prompt = basePT
+    try:
+        prompt = request.prompt
+    except Exception as e:
+        prompt = basePT
+
+    return pipeline(
+        prompt,
+        generator=Generator(pipeline.device).manual_seed(request.seed),
+        guidance_scale=0.0,
+        num_inference_steps=4,
+        max_sequence_length=256,
+        height=request.height,
+        width=request.width,
+    ).images[0]