|
|
|
|
|
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 |
|
|
|
|
|
os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True" |
|
|
os.environ["TOKENIZERS_PARALLELISM"] = "True" |
|
|
torch._dynamo.config.suppress_errors = True |
|
|
|
|
|
Pipeline = None |
|
|
ids = "black-forest-labs/FLUX.1-schnell" |
|
|
Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9" |
|
|
ckpt_id = "agentbot/Quant-Flux-2" |
|
|
ckpt_revision = "2996318a090bd3bd4b079413d80bbcfd95e6febc" |
|
|
|
|
|
def load_pipeline() -> Pipeline: |
|
|
vae = AutoencoderKL.from_pretrained(ids,revision=Revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,) |
|
|
quantize_(vae, int8_weight_only()) |
|
|
|
|
|
|
|
|
text_encoder_2 = T5EncoderModel.from_pretrained("agentbot/t5-v1_1-xxl-encoder-bf16_", revision = "208e3686b3027985dbd8c9098c273e0155c77ef4", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last) |
|
|
path = os.path.join(HF_HUB_CACHE, "models--agentbot--FLUX.1-schnell-int8wo_/snapshots/aa66177be06aba5a88dbe7265255bec48833a936") |
|
|
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, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,) |
|
|
pipeline.to("cuda") |
|
|
|
|
|
try: |
|
|
flux_transformer = FluxTransformer2DModel.from_pretrained(flux_path, torch_dtype=torch.bfloat16, use_safetensors=False) |
|
|
flux_pipeline = FluxPipeline.from_pretrained(ckpt_id, revision=ckpt_revision, transformer=flux_transformer, local_files_only=True, torch_dtype=torch.bfloat16) |
|
|
flux_pipeline.to("cuda") |
|
|
with torch.inference_mode(): |
|
|
flux_pipeline(prompt="satiety, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256) |
|
|
except Exception as e: |
|
|
flux_pipeline = None |
|
|
|
|
|
if flux_pipeline: |
|
|
return flux_pipeline |
|
|
|
|
|
for _ in range(2): |
|
|
pipeline(prompt="satiety, unwitherable, Pygmy, ramlike, Curtis, fingerstone, rewhisper", 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] |
|
|
|
|
|
|
