hoelee/stable-diffusion-webui
1
0
Fork 0
mirror of https://github.com/AUTOMATIC1111/stable-diffusion-webui.git synced 2025-08-04 11:12:35 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Add CPU fp8 support

Browse Source 
Since norm layer need fp32, I only convert the linear operation layer(conv2d/linear)

And TE have some pytorch function not support bf16 amp in CPU. I add a condition to indicate if the autocast is for unet.
This commit is contained in:
Kohaku-Blueleaf
2023-10-24 01:49:05 +08:00
parent 5f9ddfa46f
commit eaa9f5162f
3 changed files with 22 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
modules/sd_models.py
View File

@@ -391,12 +391,24 @@ 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:
if shared.cmd_opts.opt_unet_fp8_storage:
enable_fp8 = True
elif model.is_sdxl and shared.cmd_opts.opt_unet_fp8_storage_xl:
enable_fp8 = True
if enable_fp8:
devices.fp8 = True
if devices.device == devices.cpu:
for module in model.model.diffusion_model.modules():
if isinstance(module, torch.nn.Conv2d):
module.to(torch.float8_e4m3fn)
elif isinstance(module, torch.nn.Linear):
module.to(torch.float8_e4m3fn)
timer.record("apply fp8 unet for cpu")
else:
model.model.diffusion_model = model.model.diffusion_model.to(torch.float8_e4m3fn)
timer.record("apply fp8 unet")
elif model.is_sdxl and shared.cmd_opts.opt_unet_fp8_storage_xl:
model.model.diffusion_model = model.model.diffusion_model.to(torch.float8_e4m3fn)
timer.record("apply fp8 unet for sdxl")
devices.unet_needs_upcast = shared.cmd_opts.upcast_sampling and devices.dtype == torch.float16 and devices.dtype_unet == torch.float16