AUTOMATIC1111/stable-diffusion-webui
1
0
Fork 0
mirror of https://github.com/AUTOMATIC1111/stable-diffusion-webui.git synced 2026-03-06 13:40:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add fp8 for sd unet

Browse Source
This commit is contained in:
Kohaku-Blueleaf
2023-10-19 13:56:17 +08:00
parent 861cbd5636
commit 7c128bbdac
11 changed files with 36 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
modules/cmd_args.py
View File

@@ -118,3 +118,4 @@ parser.add_argument('--timeout-keep-alive', type=int, default=30, help='set time
parser.add_argument("--disable-all-extensions", action='store_true', help="prevent all extensions from running regardless of any other settings", default=False)
parser.add_argument("--disable-extra-extensions", action='store_true', help="prevent all extensions except built-in from running regardless of any other settings", default=False)
parser.add_argument("--skip-load-model-at-start", action='store_true', help="if load a model at web start, only take effect when --nowebui", )
parser.add_argument("--opt-unet-fp8-storage", action='store_true', help="use fp8 for SD UNet to save vram", default=False)

3
modules/sd_models.py
View File

@@ -391,6 +391,9 @@ def load_model_weights(model, checkpoint_info: CheckpointInfo, state_dict, timer
devices.dtype_unet = torch.float16
timer.record("apply half()")
if shared.cmd_opts.opt_unet_fp8_storage:
model.model.diffusion_model = model.model.diffusion_model.to(torch.float8_e4m3fn)
timer.record("apply fp8 unet")
devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16