| from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny |
| from diffusers.image_processor import VaeImageProcessor |
| from diffusers.schedulers import FlowMatchEulerDiscreteScheduler |
| from huggingface_hub.constants import HF_HUB_CACHE |
| 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, fpx_weight_only |
| import os |
|
|
| os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" |
|
|
| Pipeline = None |
|
|
| ckpt_id = "malicious546/flux.1-schnell-dat-q8" |
| ckpt_revision = "f312bd733a4f351e328e94f12e1cd4d0d4baa5eb" |
|
|
|
|
| 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) |
|
|
| path = os.path.join( |
| HF_HUB_CACHE, |
| "models--malicious546--flux.1-schnell-dat-int8/snapshots/017b51c93e8b1a0a7ab7a05cdcce2270f139de60", |
| ) |
| model = FluxTransformer2DModel.from_pretrained( |
| path, torch_dtype=dtype, use_safetensors=False |
| ).to(memory_format=torch.channels_last) |
| pipeline = DiffusionPipeline.from_pretrained( |
| ckpt_id, |
| revision=ckpt_revision, |
| transformer=model, |
| text_encoder_2=text_encoder_2, |
| torch_dtype=dtype, |
| ).to(device) |
| pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune") |
|
|
| for _ in range(3): |
| pipeline( |
| prompt="divination, aftermath, airy, flatworm, adjuster, fruity, dullard, presence", |
| width=1024, |
| height=1024, |
| guidance_scale=0.0, |
| num_inference_steps=4, |
| max_sequence_length=256, |
| ) |
|
|
| empty_cache() |
| return pipeline |
|
|
|
|
| @torch.no_grad() |
| def infer( |
| request: TextToImageRequest, pipeline: Pipeline, generator: Generator |
| ) -> Image: |
| return 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] |
|
|
