stabilityai

/

stable-diffusion-xl-base-1.0

like 5.42k

>>> from diffusers.models import AutoencoderKL
>>> import torch
>>> vae16 = AutoencoderKL.from_pretrained("sdxl-vae",torch_dtype=torch.float16).to('cuda')
>>> vae = AutoencoderKL.from_pretrained("sdxl-vae").to('cuda')

>>> import numpy as np
>>> from PIL import Image
>>> x = np.array(Image.open('1.png').convert('RGB'))

>>> x = x/127.5 -1.0
>>> x = torch.from_numpy(x)
>>> x = x.permute(2,0,1)
>>> x.shape
torch.Size([3, 1024, 1024])

>>> im = x.to('cuda',dtype=torch.float32)
>>> im16 = x.to('cuda',dtype=torch.float16)
>>> t16 = vae16.encode(im16).latent_dist.sample()
>>> t16
tensor([[[[nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          ...,
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan]],

         [[nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          ...,
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan]],

         [[nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          ...,
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan]],

         [[nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          ...,
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan],
          [nan, nan, nan,  ..., nan, nan, nan]]]], device='cuda:0',
       dtype=torch.float16, grad_fn=<AddBackward0>)
>>> t = vae.encode(im).latent_dist.sample()
>>> t
tensor([[[[-10.6650, -10.7863,  -8.9922,  ...,  -4.9603,  -7.7862,  -5.2480],
          [ -7.2087, -11.2612,  -7.2466,  ...,  -5.0280,  -8.4497,  -5.2412],
          [ -9.8503,  -7.0086,  -8.6645,  ...,  -6.8731,  -5.7120,  -5.2228],
          ...,
          [ -0.8512,  -2.7118,  -3.6125,  ...,  -8.5498,  -6.7364,  -4.5577],
          [ -4.7474,  -4.0266,  -2.2864,  ...,  -5.6377,  -5.6006,  -7.2353],
          [ -3.9695,  -5.3364,  -1.7182,  ...,  -4.1163,  -3.1237,  -8.5114]],

         [[  0.5291,   3.1062,   3.9364,  ...,   6.1754,   6.9985,  10.2787],
          [  4.3542,   2.2514,   4.5842,  ...,   9.9785,   7.0942,   9.3612],
          [  2.3624,   7.6000,   4.0870,  ...,   8.2662,   7.3177,   7.4536],
          ...,
          [  2.4479,   5.3318,   7.7907,  ...,   2.2813,   3.3110,   7.9058],
          [  3.1145,   7.0900,   9.5676,  ...,   4.0560,   7.3074,   5.4294],
          [  4.1593,   2.2279,   6.3064,  ...,   5.3051,   2.4846,  -0.6007]],

         [[ -5.9309,   2.5895,  -7.6263,  ...,   7.6300,   9.7816,  -3.2628],
          [  0.0919,  -4.1963,  -0.4588,  ...,   8.5846,   7.2509,   1.5191],
          [ -7.2874,   0.2553,  -0.2289,  ...,   7.3174,  12.3518,   0.0686],
          ...,
          [  7.4820,   4.8404,   2.6403,  ...,  -3.4686,  -7.8857, -10.7730],
          [ -3.9887,   5.6537,  -5.4063,  ...,  -4.4793,   0.0799,  -6.0221],
          [  5.6877,  -0.1834,   1.8418,  ...,   0.7933,   5.1394,  -2.5688]],

         [[  5.9649,   7.5648,   6.7202,  ...,   7.6067,   2.7659,   2.5190],
          [ 11.1852,   4.1965,  11.4174,  ...,   6.5834,   0.0270,  -0.6252],
          [  5.5535,   8.5816,  10.5749,  ...,  -0.3417,   3.5020,   3.8151],
          ...,
          [  7.0386,   7.9634,   8.0208,  ...,   2.0303,   6.1680,   3.0449],
          [  4.4080,   5.6513,   4.0449,  ...,   3.7297,   9.5295,   0.9136],
          [  4.0043,   1.5003,   8.5650,  ...,  10.1849,   6.6777,   1.8639]]]],
       device='cuda:0', grad_fn=<AddBackward0>)
>>> 

When I use fp16 the result is nan, if I use fp32 the result is normal

I get it. It's my problem