File size: 4,189 Bytes
ae26e7d |
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 |
import torch
from modules import devices
module_in_gpu = None
cpu = torch.device("cpu")
def send_everything_to_cpu():
global module_in_gpu
if module_in_gpu is not None:
module_in_gpu.to(cpu)
module_in_gpu = None
def setup_for_low_vram(sd_model, use_medvram):
parents = {}
def send_me_to_gpu(module, _):
"""send this module to GPU; send whatever tracked module was previous in GPU to CPU;
we add this as forward_pre_hook to a lot of modules and this way all but one of them will
be in CPU
"""
global module_in_gpu
module = parents.get(module, module)
if module_in_gpu == module:
return
if module_in_gpu is not None:
module_in_gpu.to(cpu)
module.to(devices.device)
module_in_gpu = module
# see below for register_forward_pre_hook;
# first_stage_model does not use forward(), it uses encode/decode, so register_forward_pre_hook is
# useless here, and we just replace those methods
first_stage_model = sd_model.first_stage_model
first_stage_model_encode = sd_model.first_stage_model.encode
first_stage_model_decode = sd_model.first_stage_model.decode
def first_stage_model_encode_wrap(x):
send_me_to_gpu(first_stage_model, None)
return first_stage_model_encode(x)
def first_stage_model_decode_wrap(z):
send_me_to_gpu(first_stage_model, None)
return first_stage_model_decode(z)
# for SD1, cond_stage_model is CLIP and its NN is in the tranformer frield, but for SD2, it's open clip, and it's in model field
if hasattr(sd_model.cond_stage_model, 'model'):
sd_model.cond_stage_model.transformer = sd_model.cond_stage_model.model
# remove four big modules, cond, first_stage, depth (if applicable), and unet from the model and then
# send the model to GPU. Then put modules back. the modules will be in CPU.
stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, getattr(sd_model, 'depth_model', None), sd_model.model
sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.model = None, None, None, None
sd_model.to(devices.device)
sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.model = stored
# register hooks for those the first three models
sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu)
sd_model.first_stage_model.register_forward_pre_hook(send_me_to_gpu)
sd_model.first_stage_model.encode = first_stage_model_encode_wrap
sd_model.first_stage_model.decode = first_stage_model_decode_wrap
if sd_model.depth_model:
sd_model.depth_model.register_forward_pre_hook(send_me_to_gpu)
parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
if hasattr(sd_model.cond_stage_model, 'model'):
sd_model.cond_stage_model.model = sd_model.cond_stage_model.transformer
del sd_model.cond_stage_model.transformer
if use_medvram:
sd_model.model.register_forward_pre_hook(send_me_to_gpu)
else:
diff_model = sd_model.model.diffusion_model
# the third remaining model is still too big for 4 GB, so we also do the same for its submodules
# so that only one of them is in GPU at a time
stored = diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed
diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed = None, None, None, None
sd_model.model.to(devices.device)
diff_model.input_blocks, diff_model.middle_block, diff_model.output_blocks, diff_model.time_embed = stored
# install hooks for bits of third model
diff_model.time_embed.register_forward_pre_hook(send_me_to_gpu)
for block in diff_model.input_blocks:
block.register_forward_pre_hook(send_me_to_gpu)
diff_model.middle_block.register_forward_pre_hook(send_me_to_gpu)
for block in diff_model.output_blocks:
block.register_forward_pre_hook(send_me_to_gpu)
|