2022-10-21 16:11:07 -06:00
import sys , os , shlex
2022-10-04 03:32:22 -06:00
import contextlib
2022-09-10 23:11:27 -06:00
import torch
2022-09-12 07:34:13 -06:00
from modules import errors
2022-11-17 01:52:17 -07:00
from packaging import version
2022-09-12 07:34:13 -06:00
2022-11-12 00:00:49 -07:00
2022-11-17 01:52:17 -07:00
# has_mps is only available in nightly pytorch (for now) and macOS 12.3+.
2022-11-11 20:02:40 -07:00
# check `getattr` and try it for compatibility
def has_mps ( ) - > bool :
2022-11-12 00:00:49 -07:00
if not getattr ( torch , ' has_mps ' , False ) :
return False
2022-11-11 20:02:40 -07:00
try :
torch . zeros ( 1 ) . to ( torch . device ( " mps " ) )
return True
except Exception :
return False
2022-09-10 23:11:27 -06:00
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 ) ) :
2022-11-12 00:00:49 -07:00
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
2022-11-12 00:00:49 -07:00
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 } "
2022-10-22 05:04:14 -06:00
2022-11-27 03:08:54 -07:00
return " cuda "
2022-10-22 05:04:14 -06:00
2022-11-27 03:08:54 -07:00
2023-01-27 01:28:12 -07:00
def get_optimal_device_name ( ) :
2022-11-27 03:08:54 -07:00
if torch . cuda . is_available ( ) :
2023-01-27 01:28:12 -07:00
return get_cuda_device_string ( )
2022-09-11 09:48:36 -06:00
2022-11-11 20:02:40 -07:00
if has_mps ( ) :
2023-01-27 01:28:12 -07:00
return " mps "
return " cpu "
2022-09-11 09:48:36 -06:00
2023-01-27 01:28:12 -07:00
def get_optimal_device ( ) :
return torch . device ( get_optimal_device_name ( ) )
2022-09-11 14:24:24 -06:00
2022-12-03 08:06:33 -07:00
def get_device_for ( task ) :
from modules import shared
if task in shared . cmd_opts . use_cpu :
return cpu
return get_optimal_device ( )
2022-09-11 14:24:24 -06:00
def torch_gc ( ) :
if torch . cuda . is_available ( ) :
2022-11-27 03:08:54 -07:00
with torch . cuda . device ( get_cuda_device_string ( ) ) :
2022-11-26 16:25:16 -07:00
torch . cuda . empty_cache ( )
torch . cuda . ipc_collect ( )
2022-09-12 07:34:13 -06:00
def enable_tf32 ( ) :
if torch . cuda . is_available ( ) :
2022-12-03 06:01:23 -07:00
# 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
2022-12-03 05:57:52 -07:00
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
2022-12-03 05:57:52 -07:00
2022-09-12 07:34:13 -06:00
torch . backends . cuda . matmul . allow_tf32 = True
torch . backends . cudnn . allow_tf32 = True
2022-11-06 18:05:51 -07:00
2022-09-12 07:34:13 -06:00
errors . run ( enable_tf32 , " Enabling TF32 " )
2022-09-12 11:09:32 -06:00
2022-11-12 00:00:49 -07:00
cpu = torch . device ( " cpu " )
2022-12-03 08:06:33 -07:00
device = device_interrogate = device_gfpgan = device_esrgan = device_codeformer = None
2022-10-02 06:03:39 -06:00
dtype = torch . float16
2022-10-10 07:11:14 -06:00
dtype_vae = torch . float16
2023-01-24 21:51:45 -07:00
dtype_unet = torch . float16
unet_needs_upcast = False
2022-09-12 11:09:32 -06:00
2022-11-12 00:00:49 -07:00
2022-09-12 11:09:32 -06:00
def randn ( seed , shape ) :
torch . manual_seed ( seed )
2022-11-30 06:02:39 -07:00
if 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 )
2022-09-13 12:49:58 -06:00
def randn_without_seed ( shape ) :
if device . type == ' mps ' :
2022-11-30 06:02:39 -07:00
return torch . randn ( shape , device = cpu ) . to ( device )
2022-09-13 12:49:58 -06:00
return torch . randn ( shape , device = device )
2022-10-04 03:32:22 -06:00
2022-10-10 07:11:14 -06:00
def autocast ( disable = False ) :
2022-10-04 03:32:22 -06:00
from modules import shared
2022-10-10 07:11:14 -06:00
if disable :
return contextlib . nullcontext ( )
2022-10-04 03:32:22 -06:00
if dtype == torch . float32 or shared . cmd_opts . precision == " full " :
return contextlib . nullcontext ( )
return torch . autocast ( " cuda " )
2022-10-25 00:01:57 -06:00
2022-11-12 00:00:49 -07:00
2023-01-24 22:23:10 -07:00
def without_autocast ( disable = False ) :
return torch . autocast ( " cuda " , enabled = False ) if torch . is_autocast_enabled ( ) and not disable else contextlib . nullcontext ( )
2023-01-16 12:59:46 -07:00
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
2023-01-16 12:59:46 -07:00
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 :
2023-01-24 22:23:10 -07:00
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. "
2023-01-16 12:59:46 -07:00
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. "
2023-01-27 03:08:00 -07:00
message + = " Use --disable-nan-check commandline argument to disable this check. "
2023-01-16 12:59:46 -07:00
raise NansException ( message )
2022-10-25 00:01:57 -06:00
# MPS workaround for https://github.com/pytorch/pytorch/issues/79383
2022-11-17 01:52:17 -07:00
orig_tensor_to = torch . Tensor . to
def tensor_to_fix ( self , * args , * * kwargs ) :
if self . device . type != ' mps ' and \
( ( len ( args ) > 0 and isinstance ( args [ 0 ] , torch . device ) and args [ 0 ] . type == ' mps ' ) or \
( isinstance ( kwargs . get ( ' device ' ) , torch . device ) and kwargs [ ' device ' ] . type == ' mps ' ) ) :
self = self . contiguous ( )
return orig_tensor_to ( self , * args , * * kwargs )
# MPS workaround for https://github.com/pytorch/pytorch/issues/80800
orig_layer_norm = torch . nn . functional . layer_norm
def layer_norm_fix ( * args , * * kwargs ) :
if len ( args ) > 0 and isinstance ( args [ 0 ] , torch . Tensor ) and args [ 0 ] . device . type == ' mps ' :
args = list ( args )
args [ 0 ] = args [ 0 ] . contiguous ( )
return orig_layer_norm ( * args , * * kwargs )
2022-12-17 01:21:19 -07:00
# MPS workaround for https://github.com/pytorch/pytorch/issues/90532
orig_tensor_numpy = torch . Tensor . numpy
def numpy_fix ( self , * args , * * kwargs ) :
if self . requires_grad :
self = self . detach ( )
return orig_tensor_numpy ( self , * args , * * kwargs )
2023-01-03 18:43:05 -07:00
# MPS workaround for https://github.com/pytorch/pytorch/issues/89784
orig_cumsum = torch . cumsum
orig_Tensor_cumsum = torch . Tensor . cumsum
def cumsum_fix ( input , cumsum_func , * args , * * kwargs ) :
if input . device . type == ' mps ' :
output_dtype = kwargs . get ( ' dtype ' , input . dtype )
2023-01-12 06:00:38 -07:00
if output_dtype == torch . int64 :
2023-01-03 18:43:05 -07:00
return cumsum_func ( input . cpu ( ) , * args , * * kwargs ) . to ( input . device )
2023-01-12 06:00:38 -07:00
elif cumsum_needs_bool_fix and output_dtype == torch . bool or cumsum_needs_int_fix and ( output_dtype == torch . int8 or output_dtype == torch . int16 ) :
return cumsum_func ( input . to ( torch . int32 ) , * args , * * kwargs ) . to ( torch . int64 )
2023-01-03 18:43:05 -07:00
return cumsum_func ( input , * args , * * kwargs )
if has_mps ( ) :
if version . parse ( torch . __version__ ) < version . parse ( " 1.13 " ) :
# PyTorch 1.13 doesn't need these fixes but unfortunately is slower and has regressions that prevent training from working
torch . Tensor . to = tensor_to_fix
torch . nn . functional . layer_norm = layer_norm_fix
torch . Tensor . numpy = numpy_fix
elif version . parse ( torch . __version__ ) > version . parse ( " 1.13.1 " ) :
2023-01-12 06:00:38 -07:00
cumsum_needs_int_fix = not torch . Tensor ( [ 1 , 2 ] ) . to ( torch . device ( " mps " ) ) . equal ( torch . ShortTensor ( [ 1 , 1 ] ) . to ( torch . device ( " mps " ) ) . cumsum ( 0 ) )
cumsum_needs_bool_fix = not torch . BoolTensor ( [ True , True ] ) . to ( device = torch . device ( " mps " ) , dtype = torch . int64 ) . equal ( torch . BoolTensor ( [ True , False ] ) . to ( torch . device ( " mps " ) ) . cumsum ( 0 ) )
torch . cumsum = lambda input , * args , * * kwargs : ( cumsum_fix ( input , orig_cumsum , * args , * * kwargs ) )
torch . Tensor . cumsum = lambda self , * args , * * kwargs : ( cumsum_fix ( self , orig_Tensor_cumsum , * args , * * kwargs ) )
2023-01-03 18:43:05 -07:00
orig_narrow = torch . narrow
torch . narrow = lambda * args , * * kwargs : ( orig_narrow ( * args , * * kwargs ) . clone ( ) )
2023-01-16 12:59:46 -07:00