mirror of
https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
synced 2024-12-21 07:30:02 +08:00
77 lines
3.4 KiB
Python
77 lines
3.4 KiB
Python
import torch
|
|
|
|
module_in_gpu = None
|
|
cpu = torch.device("cpu")
|
|
if torch.has_cuda:
|
|
device = gpu = torch.device("cuda")
|
|
elif torch.has_mps:
|
|
device = gpu = torch.device("mps")
|
|
else:
|
|
device = gpu = torch.device("cpu")
|
|
|
|
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(gpu)
|
|
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
|
|
def first_stage_model_encode_wrap(self, encoder, x):
|
|
send_me_to_gpu(self, None)
|
|
return encoder(x)
|
|
|
|
def first_stage_model_decode_wrap(self, decoder, z):
|
|
send_me_to_gpu(self, None)
|
|
return decoder(z)
|
|
|
|
# remove three big modules, cond, first_stage, 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, sd_model.model
|
|
sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model = None, None, None
|
|
sd_model.to(device)
|
|
sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model = stored
|
|
|
|
# register hooks for those the first two 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 = lambda x, en=sd_model.first_stage_model.encode: first_stage_model_encode_wrap(sd_model.first_stage_model, en, x)
|
|
sd_model.first_stage_model.decode = lambda z, de=sd_model.first_stage_model.decode: first_stage_model_decode_wrap(sd_model.first_stage_model, de, z)
|
|
parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
|
|
|
|
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(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)
|