wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-12-16 01:37:04 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

sd: revise hy VAE VRAM (#11105)

Browse Source 
This was recently collapsed down to rolling VAE through temporal. Clamp
The time dimension.
This commit is contained in:
rattus 2025-12-05 03:50:04 +10:00 committed by GitHub
parent 6be85c7920
commit f4bdf5f830
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
comfy/sd.py
View File

@ -483,8 +483,10 @@ class VAE:
self.latent_dim = 3
self.latent_channels = ddconfig['z_channels'] = sd["decoder.conv_in.conv.weight"].shape[1]
self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
self.memory_used_decode = lambda shape, dtype: (1500 * shape[2] * shape[3] * shape[4] * (4 * 8 * 8)) * model_management.dtype_size(dtype)
self.memory_used_encode = lambda shape, dtype: (900 * max(shape[2], 2) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
#This is likely to significantly over-estimate with single image or low frame counts as the
#implementation is able to completely skip caching. Rework if used as an image only VAE
self.memory_used_decode = lambda shape, dtype: (2800 * min(8, ((shape[2] - 1) * 4) + 1) * shape[3] * shape[4] * (8 * 8)) * model_management.dtype_size(dtype)
self.memory_used_encode = lambda shape, dtype: (1400 * min(9, shape[2]) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
elif "decoder.unpatcher3d.wavelets" in sd:
self.upscale_ratio = (lambda a: max(0, a * 8 - 7), 8, 8)