Spaces:
Runtime error
Runtime error
File size: 6,002 Bytes
f1f9265 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 |
# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# SPDX-License-Identifier: Apache-2.0
import torch
from diffusers import AutoencoderDC
from diffusers.models import AutoencoderKL
from mmcv import Registry
from termcolor import colored
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, T5EncoderModel, T5Tokenizer
from transformers import logging as transformers_logging
from diffusion.model.dc_ae.efficientvit.ae_model_zoo import DCAE_HF
from diffusion.model.utils import set_fp32_attention, set_grad_checkpoint
MODELS = Registry("models")
transformers_logging.set_verbosity_error()
def build_model(cfg, use_grad_checkpoint=False, use_fp32_attention=False, gc_step=1, **kwargs):
if isinstance(cfg, str):
cfg = dict(type=cfg)
model = MODELS.build(cfg, default_args=kwargs)
if use_grad_checkpoint:
set_grad_checkpoint(model, gc_step=gc_step)
if use_fp32_attention:
set_fp32_attention(model)
return model
def get_tokenizer_and_text_encoder(name="T5", device="cuda"):
text_encoder_dict = {
"T5": "DeepFloyd/t5-v1_1-xxl",
"T5-small": "google/t5-v1_1-small",
"T5-base": "google/t5-v1_1-base",
"T5-large": "google/t5-v1_1-large",
"T5-xl": "google/t5-v1_1-xl",
"T5-xxl": "google/t5-v1_1-xxl",
"gemma-2b": "google/gemma-2b",
"gemma-2b-it": "google/gemma-2b-it",
"gemma-2-2b": "google/gemma-2-2b",
"gemma-2-2b-it": "google/gemma-2-2b-it",
"gemma-2-9b": "google/gemma-2-9b",
"gemma-2-9b-it": "google/gemma-2-9b-it",
"Qwen2-0.5B-Instruct": "Qwen/Qwen2-0.5B-Instruct",
"Qwen2-1.5B-Instruct": "Qwen/Qwen2-1.5B-Instruct",
}
assert name in list(text_encoder_dict.keys()), f"not support this text encoder: {name}"
if "T5" in name:
tokenizer = T5Tokenizer.from_pretrained(text_encoder_dict[name])
text_encoder = T5EncoderModel.from_pretrained(text_encoder_dict[name], torch_dtype=torch.float16).to(device)
elif "gemma" in name or "Qwen" in name:
tokenizer = AutoTokenizer.from_pretrained(text_encoder_dict[name])
tokenizer.padding_side = "right"
text_encoder = (
AutoModelForCausalLM.from_pretrained(text_encoder_dict[name], torch_dtype=torch.bfloat16)
.get_decoder()
.to(device)
)
else:
print("error load text encoder")
exit()
return tokenizer, text_encoder
def get_vae(name, model_path, device="cuda"):
if name == "sdxl" or name == "sd3":
vae = AutoencoderKL.from_pretrained(model_path).to(device).to(torch.float16)
if name == "sdxl":
vae.config.shift_factor = 0
return vae
elif "dc-ae" in name:
print(colored(f"[DC-AE] Loading model from {model_path}", attrs=["bold"]))
dc_ae = DCAE_HF.from_pretrained(model_path).to(device).eval()
return dc_ae
elif "AutoencoderDC" in name:
print(colored(f"[AutoencoderDC] Loading model from {model_path}", attrs=["bold"]))
dc_ae = AutoencoderDC.from_pretrained(model_path).to(device).eval()
return dc_ae
else:
print("error load vae")
exit()
def vae_encode(name, vae, images, sample_posterior, device):
if name == "sdxl" or name == "sd3":
posterior = vae.encode(images.to(device)).latent_dist
if sample_posterior:
z = posterior.sample()
else:
z = posterior.mode()
z = (z - vae.config.shift_factor) * vae.config.scaling_factor
elif "dc-ae" in name:
ae = vae
scaling_factor = ae.cfg.scaling_factor if ae.cfg.scaling_factor else 0.41407
z = ae.encode(images.to(device))
z = z * scaling_factor
elif "AutoencoderDC" in name:
ae = vae
scaling_factor = ae.config.scaling_factor if ae.config.scaling_factor else 0.41407
z = ae.encode(images.to(device))[0]
z = z * scaling_factor
else:
print("error load vae")
exit()
return z
def vae_decode(name, vae, latent):
if name == "sdxl" or name == "sd3":
latent = (latent.detach() / vae.config.scaling_factor) + vae.config.shift_factor
samples = vae.decode(latent).sample
elif "dc-ae" in name:
ae = vae
vae_scale_factor = (
2 ** (len(ae.config.encoder_block_out_channels) - 1)
if hasattr(ae, "config") and ae.config is not None
else 32
)
scaling_factor = ae.cfg.scaling_factor if ae.cfg.scaling_factor else 0.41407
if latent.shape[-1] * vae_scale_factor > 4000 or latent.shape[-2] * vae_scale_factor > 4000:
from patch_conv import convert_model
ae = convert_model(ae, splits=4)
samples = ae.decode(latent.detach() / scaling_factor)
elif "AutoencoderDC" in name:
ae = vae
scaling_factor = ae.config.scaling_factor if ae.config.scaling_factor else 0.41407
try:
samples = ae.decode(latent / scaling_factor, return_dict=False)[0]
except torch.cuda.OutOfMemoryError as e:
print("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
ae.enable_tiling(tile_sample_min_height=1024, tile_sample_min_width=1024)
samples = ae.decode(latent / scaling_factor, return_dict=False)[0]
else:
print("error load vae")
exit()
return samples
|