|
|
|
|
|
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, 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 |
|
|
from para_attn.first_block_cache.diffusers_adapters import apply_cache_on_pipe |
|
|
os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True" |
|
|
os.environ["TOKENIZERS_PARALLELISM"] = "True" |
|
|
torch._dynamo.config.suppress_errors = True |
|
|
Pipeline = None |
|
|
ids = "slobers/Flux.1.Schnella" |
|
|
Revision = "e34d670e44cecbbc90e4962e7aada2ac5ce8b55b" |
|
|
gc.collect() |
|
|
torch.cuda.empty_cache() |
|
|
torch.cuda.reset_max_memory_allocated() |
|
|
torch.cuda.reset_peak_memory_stats() |
|
|
def load_pipeline() -> Pipeline: |
|
|
path = os.path.join(HF_HUB_CACHE, "models--slobers--Flux.1.Schnella/snapshots/e34d670e44cecbbc90e4962e7aada2ac5ce8b55b/transformer") |
|
|
transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last) |
|
|
pipeline = FluxPipeline.from_pretrained(ids, revision=Revision, transformer=transformer, local_files_only=True, torch_dtype=torch.bfloat16,) |
|
|
pipeline.to("cuda") |
|
|
quantize_(pipeline.vae, int8_weight_only()) |
|
|
apply_cache_on_pipe(pipeline, residual_diff_threshold=0.557) |
|
|
|
|
|
pipeline.vae = torch.compile(pipeline.vae, mode="max-autotune", fullgraph=True) |
|
|
for _ in range(2): |
|
|
pipeline(prompt="insensible, timbale, pothery, electrovital, actinogram, boobs, intracerebellar, centrodesmus", 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) -> Image: |
|
|
generator = Generator(pipeline.device).manual_seed(request.seed) |
|
|
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] |
