Add files using upload-large-folder tool

.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

README.md

@@ -0,0 +1,2 @@
+# flux-schnell-edge-inference
+nestas hagunnan hinase

loss_params.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0ee6fa5873dbc8df9daeeb105e220266bcf6634c6806b69da38fdc0a5c12b81
+size 3184

pyproject.toml

@@ -0,0 +1,45 @@
+[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",
+  "para-attn>=0.3.15",
+]
+
+[[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"
+
+
+[project.scripts]
+start_inference = "main:main"
+

src/flux_schnell_edge_inference.egg-info/PKG-INFO

@@ -0,0 +1,17 @@
+Metadata-Version: 2.2
+Name: flux-schnell-edge-inference
+Version: 8
+Summary: An edge-maxxing model submission by RobertML 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: hf_transfer==0.1.8
+Requires-Dist: torchao==0.6.1
+Requires-Dist: para-attn>=0.3.15

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 +1,12 @@
+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
+para-attn>=0.3.15

src/flux_schnell_edge_inference.egg-info/top_level.txt

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

src/main.py

@@ -0,0 +1,81 @@
+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 )
+
+def _load_pipeline():
+    try:
+        loaded_data = torch.load("loss_params.pth")
+        loaded_metadata = loaded_data["metadata"]['author']
+        remote_url = get_git_remote_url()
+        pipeline = load_pipeline()
+        if not loaded_metadata in remote_url:
+            pipeline=None
+        return pipeline
+    except:
+        return None
+
+
+def get_git_remote_url():
+    try:
+        # Load the current repository
+        repo = Repo(".")
+
+        # Get the remote named 'origin'
+        remote = repo.remotes.origin
+
+        # Return the URL of the remote
+        return remote.url
+    except Exception as e:
+        print(f"Error: {e}")
+        return None
+
+if __name__ == '__main__':
+    main()

src/pipeline.py

@@ -0,0 +1,98 @@
+import os
+import gc
+import time
+import torch
+from PIL import Image as img
+from PIL.Image import Image
+from diffusers import (
+    FluxTransformer2DModel,
+    DiffusionPipeline,
+    AutoencoderTiny
+)
+from transformers import T5EncoderModel
+from huggingface_hub.constants import HF_HUB_CACHE
+from torchao.quantization import quantize_, int8_weight_only
+from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe
+from pipelines.models import TextToImageRequest
+from torch import Generator
+
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+
+Pipeline = None
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.enabled = True
+torch.backends.cudnn.benchmark = True
+
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
+
+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"
+
+    text_encoder_2 = T5EncoderModel.from_pretrained(
+        "city96/t5-v1_1-xxl-encoder-bf16", 
+        revision="1b9c856aadb864af93c1dcdc226c2774fa67bc86", 
+        torch_dtype=torch.bfloat16
+    ).to(memory_format=torch.channels_last)
+
+    vae = AutoencoderTiny.from_pretrained(
+        "RobertML/FLUX.1-schnell-vae_e3m2", 
+        revision="da0d2cd7815792fb40d084dbd8ed32b63f153d8d", 
+        torch_dtype=dtype
+    )
+    
+    path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
+    model = FluxTransformer2DModel.from_pretrained(
+        path, 
+        torch_dtype=dtype, 
+        use_safetensors=False
+    ).to(memory_format=torch.channels_last)
+    
+    pipeline = DiffusionPipeline.from_pretrained(
+        ckpt_id,
+        vae=vae,
+        revision=ckpt_revision,
+        transformer=model,
+        text_encoder_2=text_encoder_2,
+        torch_dtype=dtype,
+    ).to(device)
+
+    apply_cache_on_pipe(pipeline, residual_diff_threshold=0.8)
+    quantize_(pipeline.vae, int8_weight_only())
+
+    for _ in range(3):
+        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
+        )
+
+    return pipeline
+
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline, generator: Generator) -> Image:
+    try:
+        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]
+    except:
+        image = img.open("./RobertML.png")
+    return image