princeDjoumessi/CogVideoX-5b-nf4

CogVideoX-5b NF4 (4-bit quantized)

Version quantifiée NF4 de THUDM/CogVideoX-5b-I2V pour réduire l'empreinte VRAM (~50% vs bfloat16).

Utilisation

from diffusers import CogVideoXImageToVideoPipeline, CogVideoXTransformer3DModel
from transformers import T5EncoderModel
import torch

transformer = CogVideoXTransformer3DModel.from_pretrained(
    'princeDjoumessi/CogVideoX-5b-nf4',
    subfolder='transformer',
    torch_dtype=torch.bfloat16,
)
text_encoder = T5EncoderModel.from_pretrained(
    'princeDjoumessi/CogVideoX-5b-nf4',
    subfolder='text_encoder',
    torch_dtype=torch.bfloat16,
)
pipe = CogVideoXImageToVideoPipeline.from_pretrained(
    'THUDM/CogVideoX-5b-I2V',
    transformer=transformer,
    text_encoder=text_encoder,
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
pipe.vae.enable_slicing()

Composants quantifiés

Composant	Quantification
Transformer3D	NF4 (double quant)
Text Encoder (T5)	NF4 (double quant)
VAE	bfloat16 (inchangé)
Scheduler	inchangé

Généré sur Kaggle T4 avec bitsandbytes.