dartrs

Warning

wip

Python

pip install -U dartrs

from dartrs.dartrs import DartTokenizer
from dartrs.utils import get_generation_config
from dartrs.v2 import (
    compose_prompt,
    MixtralModel,
    V2Model,
)
import time
import os

MODEL_NAME = "p1atdev/dart-v2-moe-sft"

model = MixtralModel.from_pretrained(MODEL_NAME, dtype="fp16")
tokenizer = DartTokenizer.from_pretrained(MODEL_NAME)

config = get_generation_config(
    prompt=compose_prompt(
        copyright="vocaloid",
        character="hatsune miku",
        rating="general", # sfw, general, sensitive, nsfw, questionable, explicit
        aspect_ratio="tall", # ultra_wide, wide, square, tall, ultra_tall
        length="medium", # very_short, short, medium, long, very_long
        identity="none", # none, lax, strict
        prompt="1girl, cat ears",
    ),
    tokenizer=tokenizer,
    seed=42,
)

start = time.time()
output = model.generate(config)
end = time.time()

print(output)
print(f"Time taken: {end - start:.2f}s")
# Output:
# cowboy shot, detached sleeves, expressionless, from side, hair ornament, halftone, hairclip, holding, limited palette, long hair, looking at viewer, miniskirt, necktie, pleated skirt, shirt, simple background, skirt, sleeveless, sleeveless shirt, straight-on, thighhighs, twintails, very long hair, zettai ryouiki
# Time taken: 0.22s