Update pyproject.toml

pyproject.toml

@@ -1,42 +1,47 @@
-[build-system]
-requires = ["setuptools >= 75.0"]
-build-backend = "setuptools.build_meta"
+#6
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import torch._dynamo
+import gc
+import os
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
+from PIL.Image import Image
+from pipelines.models import TextToImageRequest
+from torch import Generator
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
 
-[project]
-name = "flux-schnell-edge-inference"
-description = "An edge-maxxing model submission 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",
-  "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",
-]
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+torch._dynamo.config.suppress_errors = True
 
-[[tool.edge-maxxing.models]]
-repository = "black-forest-labs/FLUX.1-schnell"
-revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
-exclude = ["transformer", "vae", "text_encoder_2"]
+Pipeline = None
+ids = "black-forest-labs/FLUX.1-schnell"
+Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
 
-[[tool.edge-maxxing.models]]
-repository = "RobertML/FLUX.1-schnell-int8wo"
-revision = "307e0777d92df966a3c0f99f31a6ee8957a9857a"
+def load_pipeline() -> Pipeline:
+    vae = AutoencoderTiny.from_pretrained("slobers/tt1",revision="ec746bf42d91e3335760895281f070df54f2196a", torch_dtype=torch.bfloat16,)
+    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)
+    path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
+    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
+    pipeline = DiffusionPipeline.from_pretrained(ids, revision=Revision, vae=vae, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
+    pipeline.to("cuda")
 
-[[tool.edge-maxxing.models]]
-repository = "city96/t5-v1_1-xxl-encoder-bf16"
-revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86"
+    for _ in range(3):
+        pipeline(prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    return pipeline
 
-[[tool.edge-maxxing.models]]
-repository = "slobers/tt1"
-revision = "ec746bf42d91e3335760895281f070df54f2196a"
+@torch.no_grad()
+def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    generator = Generator(pipeline.device).manual_seed(request.seed)
 
-[project.scripts]
-start_inference = "main:main"
+    return pipeline(
+        request.prompt,
+        generator=generator,
+        guidance_scale=0.0,
+        num_inference_steps=4,
+        max_sequence_length=256,
+        height=request.height,
+        width=request.width,
+    ).images[0]