aboutsummaryrefslogtreecommitdiffstats
path: root/modules
diff options
context:
space:
mode:
authorKohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com>2023-10-19 05:56:17 +0000
committerKohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com>2023-10-19 05:56:17 +0000
commit7c128bbdac0da1767c239174e91af6f327845372 (patch)
treec7c8b1271c17bab0f5e71ea1d97a21bbbc85332e /modules
parent861cbd56363ffa0df3351cf1162f507425a178cd (diff)
downloadstable-diffusion-webui-gfx803-7c128bbdac0da1767c239174e91af6f327845372.tar.gz
stable-diffusion-webui-gfx803-7c128bbdac0da1767c239174e91af6f327845372.tar.bz2
stable-diffusion-webui-gfx803-7c128bbdac0da1767c239174e91af6f327845372.zip
Add fp8 for sd unet
Diffstat (limited to 'modules')
-rw-r--r--modules/cmd_args.py1
-rw-r--r--modules/sd_models.py3
2 files changed, 4 insertions, 0 deletions
diff --git a/modules/cmd_args.py b/modules/cmd_args.py
index 4e602a84..0f14c71e 100644
--- a/modules/cmd_args.py
+++ b/modules/cmd_args.py
@@ -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)
diff --git a/modules/sd_models.py b/modules/sd_models.py
index 3b6cdea1..3b8ff820 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -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