stable-diffusion-webui/modules/devices.py

157 lines
4.1 KiB
Python
Raw Normal View History

import sys
import contextlib
2022-09-10 23:11:27 -06:00
import torch
from modules import errors
if sys.platform == "darwin":
from modules import mac_specific
def has_mps() -> bool:
if sys.platform != "darwin":
return False
else:
return mac_specific.has_mps
2022-09-11 09:48:36 -06:00
2022-10-21 16:11:07 -06:00
def extract_device_id(args, name):
for x in range(len(args)):
if name in args[x]:
return args[x + 1]
2022-10-21 16:11:07 -06:00
return None
2022-09-11 09:48:36 -06:00
2022-11-27 03:08:54 -07:00
def get_cuda_device_string():
from modules import shared
if shared.cmd_opts.device_id is not None:
return f"cuda:{shared.cmd_opts.device_id}"
2022-11-27 03:08:54 -07:00
return "cuda"
2022-11-27 03:08:54 -07:00
def get_optimal_device_name():
2022-11-27 03:08:54 -07:00
if torch.cuda.is_available():
return get_cuda_device_string()
2022-09-11 09:48:36 -06:00
if has_mps():
return "mps"
return "cpu"
2022-09-11 09:48:36 -06:00
def get_optimal_device():
return torch.device(get_optimal_device_name())
def get_device_for(task):
from modules import shared
if task in shared.cmd_opts.use_cpu:
return cpu
return get_optimal_device()
def torch_gc():
if torch.cuda.is_available():
2022-11-27 03:08:54 -07:00
with torch.cuda.device(get_cuda_device_string()):
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
def enable_tf32():
if torch.cuda.is_available():
# enabling benchmark option seems to enable a range of cards to do fp16 when they otherwise can't
# see https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/4407
2023-05-10 02:05:02 -06:00
if any(torch.cuda.get_device_capability(devid) == (7, 5) for devid in range(0, torch.cuda.device_count())):
2022-11-07 19:06:48 -07:00
torch.backends.cudnn.benchmark = True
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
2022-11-06 18:05:51 -07:00
errors.run(enable_tf32, "Enabling TF32")
2022-09-12 11:09:32 -06:00
cpu = torch.device("cpu")
device = device_interrogate = device_gfpgan = device_esrgan = device_codeformer = None
dtype = torch.float16
2022-10-10 07:11:14 -06:00
dtype_vae = torch.float16
dtype_unet = torch.float16
unet_needs_upcast = False
2022-09-12 11:09:32 -06:00
def cond_cast_unet(input):
return input.to(dtype_unet) if unet_needs_upcast else input
def cond_cast_float(input):
return input.float() if unet_needs_upcast else input
2022-09-12 11:09:32 -06:00
def randn(seed, shape):
from modules.shared import opts
2022-09-12 11:09:32 -06:00
torch.manual_seed(seed)
if opts.randn_source == "CPU" or device.type == 'mps':
return torch.randn(shape, device=cpu).to(device)
2022-09-12 11:09:32 -06:00
return torch.randn(shape, device=device)
def randn_without_seed(shape):
from modules.shared import opts
if opts.randn_source == "CPU" or device.type == 'mps':
return torch.randn(shape, device=cpu).to(device)
return torch.randn(shape, device=device)
2022-10-10 07:11:14 -06:00
def autocast(disable=False):
from modules import shared
2022-10-10 07:11:14 -06:00
if disable:
return contextlib.nullcontext()
if dtype == torch.float32 or shared.cmd_opts.precision == "full":
return contextlib.nullcontext()
return torch.autocast("cuda")
def without_autocast(disable=False):
return torch.autocast("cuda", enabled=False) if torch.is_autocast_enabled() and not disable else contextlib.nullcontext()
class NansException(Exception):
pass
def test_for_nans(x, where):
from modules import shared
2023-01-17 01:04:56 -07:00
if shared.cmd_opts.disable_nan_check:
return
if not torch.all(torch.isnan(x)).item():
return
if where == "unet":
message = "A tensor with all NaNs was produced in Unet."
if not shared.cmd_opts.no_half:
message += " This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the \"Upcast cross attention layer to float32\" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this."
elif where == "vae":
message = "A tensor with all NaNs was produced in VAE."
if not shared.cmd_opts.no_half and not shared.cmd_opts.no_half_vae:
message += " This could be because there's not enough precision to represent the picture. Try adding --no-half-vae commandline argument to fix this."
else:
message = "A tensor with all NaNs was produced."
message += " Use --disable-nan-check commandline argument to disable this check."
raise NansException(message)