Initial commit with folder contents

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* 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

pyproject.toml

@@ -0,0 +1,43 @@
+[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",
+  "torchao>=0.6.1",
+]
+
+[[tool.edge-maxxing.models]]
+repository = "black-forest-labs/FLUX.1-schnell"
+revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+exclude = ["transformer"]
+
+[[tool.edge-maxxing.models]]
+repository = "proact/PRO_FLUX_0load"
+revision = "22f140d15b1b2f86794cda5ffb7bc52e7b917965"
+
+[[tool.edge-maxxing.models]]
+repository = "city96/t5-v1_1-xxl-encoder-bf16"
+revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+
+[[tool.edge-maxxing.models]]
+repository = "proact/PRO_FLUX_1load"
+revision = "91741cd838292c87e50d14d1b5d14335e180a961"
+
+
+[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 +1,59 @@
+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
+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 +1,95 @@
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import torch._dynamo
+import gc
+from PIL import Image as img
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+import time
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only
+#from torchao.quantization import autoquant
+
+Pipeline = None
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.cuda.set_per_process_memory_fraction(0.95)
+
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+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()
+
+    dtype, device = torch.bfloat16, "cuda"
+
+    try:
+        vae = AutoencoderTiny.from_pretrained(
+            "/home/sandbox/.cache/huggingface/hub/models--proact--PRO_FLUX_1load/snapshots/91741cd838292c87e50d14d1b5d14335e180a961", torch_dtype=dtype)
+    except:
+        vae = AutoencoderTiny.from_pretrained("proact/PRO_FLUX_1load", torch_dtype=dtype)
+
+    try:
+        vae.enable_slicing()
+        vae.enable_tiling()
+    except:
+        pass
+
+    ############ Text Encoder ############
+    text_encoder = CLIPTextModel.from_pretrained(
+        ckpt_id, subfolder="text_encoder", torch_dtype=torch.bfloat16
+    )
+    ############ Text Encoder 2 ############
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=torch.bfloat16
+    )
+
+    model = FluxTransformer2DModel.from_pretrained(
+        "/home/sandbox/.cache/huggingface/hub/models--proact--PRO_FLUX_0load/snapshots/22f140d15b1b2f86794cda5ffb7bc52e7b917965", torch_dtype=dtype, use_safetensors=False
+    )
+
+    empty_cache()
+    pipeline = DiffusionPipeline.from_pretrained(
+        ckpt_id, 
+        transformer=model,
+        text_encoder=text_encoder,
+        text_encoder_2=text_encoder_2,
+        torch_dtype=dtype,
+        vae=vae
+        ).to(device)
+
+
+    empty_cache()
+    pipeline(prompt="untenibleness, gynecocracy, overcapitalization, demiplate, shockable", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    
+    return pipeline
+
+
+@torch.inference_mode()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    empty_cache()
+    if request.seed is None:
+        generator = None
+    else:
+        generator = Generator(device="cuda").manual_seed(request.seed)
+
+    empty_cache()
+    image = pipeline(prompt=request.prompt,
+                      width=request.width,
+                      height=request.height,
+                      guidance_scale=0.0,
+                      generator=generator,
+                      output_type="pil",
+                      max_sequence_length=256,
+                      num_inference_steps=4).images[0]
+    return image