rixty
/
KoboldAI-Client
mirror of https://github.com/KoboldAI/KoboldAI-Client.git
1
1
Fork 0
Code Issues Projects Releases Wiki Activity

Seperate Low Memory Mode 

In 1.16 we had significantly faster loading speeds because we did not do as much memory conservation, its time to give users the choice. If you want the original faster behavior and have the memory run KoboldAI as usual. Otherwise run play-lowmem.bat or aiserver.py with --lowmem. For colab this is still the default behavior to avoid breaking models that would otherwise load fine.
This commit is contained in:
henk717 2022-02-18 16:21:28 +01:00
parent 4c84d731db
commit a47e93cee7
2 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
aiserver.py
View File

@ -479,6 +479,7 @@ parser.add_argument("--colab", action='store_true', help="Optimize for Google Co
parser.add_argument("--nobreakmodel", action='store_true', help="Disables Breakmodel support completely.")
parser.add_argument("--unblock", action='store_true', default=False, help="Unblocks the KoboldAI port to be accessible from other machines without optimizing for remote play (It is recommended to use --host instead)")
parser.add_argument("--quiet", action='store_true', default=False, help="If present will suppress any story related text from showing on the console")
parser.add_argument("--lowmem", action='store_true', help="Extra Low Memory loading for the GPU, slower but memory does not peak to twice the usage")
args: argparse.Namespace = None
if(os.environ.get("KOBOLDAI_ARGS") is not None):
@ -495,6 +496,7 @@ if args.colab:
args.override_delete = True;
args.nobreakmodel = True;
args.quiet = True;
args.lowmem = True;
if args.quiet:
vars.quiet = True
@ -960,7 +962,7 @@ if(not vars.model in ["InferKit", "Colab", "OAI", "ReadOnly", "TPUMeshTransforme
@contextlib.contextmanager
def maybe_use_float16(always_use=False):
if(always_use or (vars.hascuda and (vars.usegpu or vars.breakmodel))):
if(always_use or (vars.hascuda and args.lowmem and (vars.usegpu or vars.breakmodel))):
original_dtype = torch.get_default_dtype()
torch.set_default_dtype(torch.float16)
yield True

1
play-lowmem.bat Normal file
View File

@ -0,0 +1 @@
play --lowmem %*