Download src/pipeline.py from manbeast3b/simpleflux7: direct link, hf CLI and curl.
- Browser
- Download file 2.36 kB
-
https://huggingface.co/manbeast3b/simpleflux7/resolve/main/src/pipeline.py
- Command line
-
hf download hf://manbeast3b/simpleflux7/src/pipeline.py
-
curl -L -o pipeline.py https://huggingface.co/manbeast3b/simpleflux7/resolve/main/src/pipeline.py
2.36 kB
| from diffusers import AutoencoderTiny | |
| from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel | |
| from diffusers.image_processor import VaeImageProcessor | |
| import torch | |
| import torch._dynamo | |
| import gc | |
| from PIL.Image import Image | |
| from pipelines.models import TextToImageRequest | |
| from torch import Generator | |
| from diffusers import FluxPipeline | |
| from torchao.quantization import quantize_, int8_weight_only | |
| Pipeline = None | |
| MODEL_ID = "black-forest-labs/FLUX.1-schnell" | |
| DTYPE = torch.bfloat16 | |
| def clear(): | |
| gc.collect() | |
| torch.cuda.empty_cache() | |
| torch.cuda.reset_max_memory_allocated() | |
| torch.cuda.reset_peak_memory_stats() | |
| def load_pipeline() -> Pipeline: | |
| clear() | |
| dtype, device = torch.bfloat16, "cuda" | |
| vae = AutoencoderTiny.from_pretrained("manbeast3b/quantized1", torch_dtype=DTYPE) | |
| text_encoder_2 = T5EncoderModel.from_pretrained( | |
| "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=DTYPE | |
| ) | |
| pipeline = FluxPipeline.from_pretrained(MODEL_ID,vae=vae,text_encoder_2=text_encoder_2, | |
| torch_dtype=DTYPE) | |
| torch.backends.cudnn.benchmark = True | |
| torch.backends.cuda.matmul.allow_tf32 = True | |
| torch.cuda.set_per_process_memory_fraction(0.99) | |
| pipeline.text_encoder.to(memory_format=torch.channels_last) | |
| pipeline.text_encoder_2.to(memory_format=torch.channels_last) | |
| pipeline.transformer.to(memory_format=torch.channels_last) | |
| pipeline.vae.to(memory_format=torch.channels_last) | |
| pipeline.vae = torch.compile(pipeline.vae) | |
| pipeline._exclude_from_cpu_offload = ["vae"] | |
| pipeline.enable_sequential_cpu_offload() | |
| clear() | |
| for _ in range(1): | |
| pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.1, num_inference_steps=4, max_sequence_length=256) | |
| return pipeline | |
| sample = True | |
| def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image: | |
| global sample | |
| if sample: | |
| clear() | |
| sample = None | |
| generator = Generator("cuda").manual_seed(request.seed) | |
| 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] | |
| return(image) |