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

ko.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2077712511cbeb96f4d0a6a0898b78345302ddaaf196d384f69c3d9c1adad6f9
+size 4951464

ok.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b15aacc35d11e08803e9fdf07a2eed0c7f861250352f23e91cbeda4be07ad914
+size 1800013

pyproject.toml

@@ -0,0 +1,55 @@
+[build-system]
+requires = ["setuptools >= 75.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "flux-schnell-edge-inference"
+description = "An edge-maxxing model submission by RobertML 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",
+  "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
+  "gitpython>=3.1.43",
+  "hf_transfer==0.1.8",
+  "torchao==0.6.1",
+  "setuptools>=75.3.0",
+  "torchvision"
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer"]
+
+[[tool.edge-maxxing.models]]
+repository = "RobertML/FLUX.1-schnell-int8wo"
+revision = "307e0777d92df966a3c0f99f31a6ee8957a9857a"
+
+[[tool.edge-maxxing.models]]
+repository = "city96/t5-v1_1-xxl-encoder-bf16"
+revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+
+[[tool.edge-maxxing.models]]
+repository = "RobertML/FLUX.1-schnell-vae_e3m2"
+revision = "da0d2cd7815792fb40d084dbd8ed32b63f153d8d"
+
+[[tool.edge-maxxing.models]]
+repository = "madebyollin/taef1"
+revision = "2d552378e58c9c94201075708d7de4e1163b2689"
+
+[[tool.edge-maxxing.models]]
+repository = "manbeast3b/flux.1-schnell-full1"
+revision = "cb1b599b0d712b9aab2c4df3ad27b050a27ec146"
+
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/main.py

@@ -0,0 +1,55 @@
+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")
+            generator = torch.Generator("cuda")
+            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 +1,95 @@
+from diffusers import (
+    DiffusionPipeline,
+    AutoencoderKL,
+    AutoencoderTiny,
+    FluxPipeline,
+    FluxTransformer2DModel
+)
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import (
+    T5EncoderModel,
+    T5TokenizerFast,
+    CLIPTokenizer,
+    CLIPTextModel
+)
+import torch
+import torch._dynamo
+import gc
+from PIL import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+import time
+import math
+from typing import Type, Dict, Any, Tuple, Callable, Optional, Union
+import numpy as np
+import torch.nn as nn
+import torch.nn.functional as F
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+from utils import _load
+import torchvision 
+import os
+
+# preconfigs
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.enabled = True
+torch.backends.cudnn.benchmark = True
+
+# globals
+Pipeline = None
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9" 
+TinyVAE = "madebyollin/taef1"
+TinyVAE_REV = "2d552378e58c9c94201075708d7de4e1163b2689"
+
+def empty_cache():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_max_memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+
+def filter_state_dict(model, state_dict_path):
+    global E
+    state_dict = torch.load(state_dict_path, map_location="cpu", weights_only=True)
+    prefix = 'encoder.' if type(model) == E else 'decoder.'
+    return {k.strip(prefix): v for k, v in state_dict.items() if k.strip(prefix) in model.state_dict() and v.size() == model.state_dict()[k.strip(prefix)].size()}
+    
+def load_pipeline() -> Pipeline:
+    path = os.path.join(HF_HUB_CACHE, "models--manbeast3b--flux.1-schnell-full1/snapshots/cb1b599b0d712b9aab2c4df3ad27b050a27ec146/transformer")
+    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False)
+    vae = AutoencoderTiny.from_pretrained(
+            TinyVAE, 
+            revision=TinyVAE_REV, 
+            local_files_only=True,
+            torch_dtype=torch.bfloat16)
+    vae.encoder=_load(vae.encoder, "E", dtype=torch.bfloat16); vae.decoder=_load(vae.decoder, "D", dtype=torch.bfloat16)
+    
+    pipeline = FluxPipeline.from_pretrained(ckpt_id, revision=ckpt_revision, transformer=transformer, vae=vae, local_files_only=True, torch_dtype=torch.bfloat16,)
+    pipeline.to("cuda")
+
+    # Optimize memory format
+    for component in [pipeline.text_encoder, pipeline.text_encoder_2, pipeline.transformer, pipeline.vae]:
+        component.to(memory_format=torch.channels_last)
+    
+    pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=True)
+    pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune", fullgraph=True)
+        
+    for _ in range(2):
+        pipeline(prompt="insensible, timbale, pothery, electrovital, centrodesmus", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    empty_cache()
+    return pipeline
+
+
+sample = 1
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
+    global sample
+    if not sample:
+        sample=1
+        empty_cache()
+    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="pt").images[0]
+    return torchvision.transforms.functional.to_pil_image(image.to(torch.float32).mul_(2).sub_(1))# torchvision.transforms.functional.to_pil_image(image)

src/utils.py

@@ -0,0 +1,64 @@
+A=None
+e_sd_pt="ko.pth"
+d_sd_pt="ok.pth"
+import torch as t, torch.nn as nn, torch.nn.functional as F
+def C(n_in, n_out, **kwargs):
+    return nn.Conv2d(n_in, n_out, 3, padding=1, **kwargs)
+class Clamp(nn.Module):
+    def forward(self, x):
+        return t.tanh(x / 3) * 3
+class B(nn.Module):
+    def __init__(self, n_in, n_out):
+        super().__init__()
+        self.conv = nn.Sequential(C(n_in, n_out), nn.ReLU(), C(n_out, n_out), nn.ReLU(), C(n_out, n_out))
+        self.skip = nn.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
+        self.fuse = nn.ReLU()
+    def forward(self, x):
+        return self.fuse(self.conv(x) + self.skip(x))
+def E(latent_channels=4):
+    return nn.Sequential(
+        C(3, 64), B(64, 64),
+        C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
+        C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
+        C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
+        C(64, latent_channels),
+    )
+def D(latent_channels=16):
+    return nn.Sequential(
+        Clamp(), 
+        C(latent_channels, 48),nn.ReLU(),B(48, 48), B(48, 48),
+        nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48), B(48, 48), 
+        nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48),
+        nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48), 
+        C(48, 3), 
+    )    
+class M(nn.Module):
+    lm, ls = 3, 0.5
+    def __init__(s, ep="encoder.pth", dp="decoder.pth", lc=None):
+        super().__init__()
+        if lc is None: lc = s.glc(str(ep))
+        s.e, s.d = E(lc), D(lc)
+        def f(sd, mod, pfx):
+            f_sd = {k.strip(pfx): v for k, v in sd.items() if k.strip(pfx) in mod.state_dict() and v.size() == mod.state_dict()[k.strip(pfx)].size()}
+            mod.load_state_dict(f_sd, strict=False)
+        if ep: f(t.load(ep, map_location="cpu", weights_only=True), s.e, "encoder.")
+        if dp: f(t.load(dp, map_location="cpu", weights_only=True), s.d, "decoder.")
+        s.e.requires_grad_(False)
+        s.d.requires_grad_(False)
+    def glc(s, ep): return 16 if "taef1" in ep or "taesd3" in ep else 4
+    @staticmethod
+    def sl(x): return x.div(2 * M.lm).add(M.ls).clamp(0, 1)
+    @staticmethod
+    def ul(x): return x.sub(M.ls).mul(2 * M.lm)
+    def forward(s, x, rl=False):
+        l, o = s.e(x), s.d(s.e(x))
+        return (o.clamp(0, 1), l) if rl else o.clamp(0, 1)
+def filter_state_dict(model, name):
+    state_dict = t.load(e_sd_pt if name=="E" else d_sd_pt, map_location="cpu", weights_only=True)
+    prefix = 'encoder.' if name=="E" else 'decoder.'
+    return {k.strip(prefix): v for k, v in state_dict.items() if k.strip(prefix) in model.state_dict() and v.size() == model.state_dict()[k.strip(prefix)].size()}
+def _load(model, name, dtype=t.bfloat16):
+    model = E(16) if name=="E" else D(16)
+    model.load_state_dict(filter_state_dict(model, name), strict=False)
+    model.requires_grad_(False).to(dtype=dtype)
+    return model