sample_quant.py

"""Load a 1-bit T.I.T.S.2 export and generate from it. See quantize.py for the format."""
import argparse

import torch
from safetensors.torch import load_file
from torchvision.utils import save_image

from model2 import TITS2
from preview import decode, sample_latents
from quantize import unpack
from text_encoder2 import FrozenCLIPTextEncoder


def load_1bit(path):
    raw = load_file(path)
    sd = {}
    for k, v in raw.items():
        if k.endswith(".packed"):
            base = k[: -len(".packed")]
            shape = raw[base + ".shape"].tolist()
            sd[base] = unpack(v, raw[base + ".scale"], shape[1])
        elif not k.endswith((".scale", ".shape")):
            sd[k] = v.float()
    return sd


def main():
    p = argparse.ArgumentParser()
    p.add_argument("--weights", default="exports/tits2_1bit.safetensors")
    p.add_argument("--prompts", nargs="+", default=["a red double decker bus on a city street"])
    p.add_argument("--steps", type=int, default=30)
    p.add_argument("--guidance", type=float, default=4.5)
    p.add_argument("--seed", type=int, default=7)
    p.add_argument("--out", default="out_1bit.png")
    p.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
    args = p.parse_args()

    te = FrozenCLIPTextEncoder(device=args.device)
    sd = load_1bit(args.weights) if "1bit" in args.weights else {k: v.float() for k, v in load_file(args.weights).items()}
    model = TITS2(dim=576, depth=12, heads=9, text_dim=te.embed_dim).to(args.device)
    model.load_state_dict(sd)
    model.eval()

    lat = sample_latents(model, te, args.prompts, steps=args.steps, guidance=args.guidance,
                         device=args.device, seed=args.seed)
    save_image(decode(lat, args.device), args.out, nrow=2, padding=4, pad_value=1)
    print(f"saved {args.out}")


if __name__ == "__main__":
    main()