manbeast3b commited on
Commit
58d3b2e
0 Parent(s):

Initial commit

Browse files
Files changed (5) hide show
  1. .gitattributes +37 -0
  2. pyproject.toml +45 -0
  3. src/main.py +54 -0
  4. src/pipeline.py +60 -0
  5. uv.lock +0 -0
.gitattributes ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tar filter=lfs diff=lfs merge=lfs -text
29
+ *.tflite filter=lfs diff=lfs merge=lfs -text
30
+ *.tgz filter=lfs diff=lfs merge=lfs -text
31
+ *.wasm filter=lfs diff=lfs merge=lfs -text
32
+ *.xz filter=lfs diff=lfs merge=lfs -text
33
+ *.zip filter=lfs diff=lfs merge=lfs -text
34
+ *.zst filter=lfs diff=lfs merge=lfs -text
35
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ RobertML.png filter=lfs diff=lfs merge=lfs -text
37
+ backup.png filter=lfs diff=lfs merge=lfs -text
pyproject.toml ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [build-system]
2
+ requires = ["setuptools >= 75.0"]
3
+ build-backend = "setuptools.build_meta"
4
+
5
+ [project]
6
+ name = "flux-schnell-edge-inference"
7
+ description = "An edge-maxxing model submission by RobertML for the 4090 Flux contest"
8
+ requires-python = ">=3.10,<3.13"
9
+ version = "8"
10
+ dependencies = [
11
+ "diffusers==0.31.0",
12
+ "transformers==4.46.2",
13
+ "accelerate==1.1.0",
14
+ "omegaconf==2.3.0",
15
+ "torch==2.5.1",
16
+ "protobuf==5.28.3",
17
+ "sentencepiece==0.2.0",
18
+ "edge-maxxing-pipelines @ git+https://github.com/womboai/edge-maxxing@7c760ac54f6052803dadb3ade8ebfc9679a94589#subdirectory=pipelines",
19
+ "gitpython>=3.1.43",
20
+ "hf_transfer==0.1.8",
21
+ "torchao==0.6.1",
22
+ "setuptools>=75.3.0",
23
+ ]
24
+
25
+ [[tool.edge-maxxing.models]]
26
+ repository = "black-forest-labs/FLUX.1-schnell"
27
+ revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
28
+ exclude = ["transformer"]
29
+
30
+ [[tool.edge-maxxing.models]]
31
+ repository = "RobertML/FLUX.1-schnell-int8wo"
32
+ revision = "307e0777d92df966a3c0f99f31a6ee8957a9857a"
33
+
34
+ [[tool.edge-maxxing.models]]
35
+ repository = "city96/t5-v1_1-xxl-encoder-bf16"
36
+ revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
37
+
38
+ [[tool.edge-maxxing.models]]
39
+ repository = "manbeast3b/flux.1-schnell-vae-quant1"
40
+ revision = "21552e8b4281494f0e1b969d711ec44d0a01c346"
41
+
42
+
43
+ [project.scripts]
44
+ start_inference = "main:main"
45
+
src/main.py ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import atexit
2
+ from io import BytesIO
3
+ from multiprocessing.connection import Listener
4
+ from os import chmod, remove
5
+ from os.path import abspath, exists
6
+ from pathlib import Path
7
+ from git import Repo
8
+ import torch
9
+
10
+ from PIL.JpegImagePlugin import JpegImageFile
11
+ from pipelines.models import TextToImageRequest
12
+ from pipeline import load_pipeline, infer
13
+ SOCKET = abspath(Path(__file__).parent.parent / "inferences.sock")
14
+
15
+
16
+ def at_exit():
17
+ torch.cuda.empty_cache()
18
+
19
+
20
+ def main():
21
+ atexit.register(at_exit)
22
+
23
+ print(f"Loading pipeline")
24
+ pipeline = load_pipeline()
25
+
26
+ print(f"Pipeline loaded, creating socket at '{SOCKET}'")
27
+
28
+ if exists(SOCKET):
29
+ remove(SOCKET)
30
+
31
+ with Listener(SOCKET) as listener:
32
+ chmod(SOCKET, 0o777)
33
+
34
+ print(f"Awaiting connections")
35
+ with listener.accept() as connection:
36
+ print(f"Connected")
37
+ generator = torch.Generator("cuda")
38
+ while True:
39
+ try:
40
+ request = TextToImageRequest.model_validate_json(connection.recv_bytes().decode("utf-8"))
41
+ except EOFError:
42
+ print(f"Inference socket exiting")
43
+
44
+ return
45
+ image = infer(request, pipeline, generator.manual_seed(request.seed))
46
+ data = BytesIO()
47
+ image.save(data, format=JpegImageFile.format)
48
+
49
+ packet = data.getvalue()
50
+
51
+ connection.send_bytes(packet )
52
+
53
+ if __name__ == '__main__':
54
+ main()
src/pipeline.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
2
+ from diffusers.image_processor import VaeImageProcessor
3
+ from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
4
+ from huggingface_hub.constants import HF_HUB_CACHE
5
+ from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
6
+ import torch
7
+ import torch._dynamo
8
+ import gc
9
+ from PIL import Image as img
10
+ from PIL.Image import Image
11
+ from pipelines.models import TextToImageRequest
12
+ from torch import Generator
13
+ import time
14
+ from diffusers import FluxTransformer2DModel, DiffusionPipeline
15
+ from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
16
+ import os
17
+ os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
18
+ torch._dynamo.config.suppress_errors = True
19
+
20
+ Pipeline = None
21
+
22
+ ckpt_id = "black-forest-labs/FLUX.1-schnell"
23
+ ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
24
+ def empty_cache():
25
+ gc.collect()
26
+ torch.cuda.empty_cache()
27
+ torch.cuda.reset_max_memory_allocated()
28
+ torch.cuda.reset_peak_memory_stats()
29
+
30
+ def load_pipeline() -> Pipeline:
31
+ empty_cache()
32
+
33
+ dtype, device = torch.bfloat16, "cuda"
34
+
35
+ text_encoder_2 = T5EncoderModel.from_pretrained(
36
+ "city96/t5-v1_1-xxl-encoder-bf16", revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86", torch_dtype=torch.bfloat16
37
+ ).to(memory_format=torch.channels_last)
38
+
39
+ path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
40
+ model = FluxTransformer2DModel.from_pretrained(path, torch_dtype=dtype, use_safetensors=False).to(memory_format=torch.channels_last)
41
+ pipeline = FluxPipeline.from_pretrained(
42
+ ckpt_id,
43
+ revision=ckpt_revision,
44
+ transformer=model,
45
+ text_encoder_2=text_encoder_2,
46
+ torch_dtype=dtype,
47
+ ).to(device)
48
+ pipeline.transformer = torch.compile(pipeline.transformer, mode="reduce-overhead")
49
+ quantize_(pipeline.vae, int8_weight_only())
50
+ for _ in range(5):
51
+ pipeline(prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
52
+
53
+ empty_cache()
54
+ return pipeline
55
+
56
+
57
+ @torch.no_grad()
58
+ def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
59
+ 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]
60
+ return(image)
uv.lock ADDED
The diff for this file is too large to render. See raw diff