JLNTV_IPD/CogVideo
4
0
Fork 0
mirror of https://github.com/THUDM/CogVideo.git synced 2026-05-11 18:24:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix incorrect frame padding formula in lora trainers

Browse Source 
The ncopy calculation used `latent.shape[2] % patch_size_t` which
computes the remainder rather than the number of frames needed to
reach alignment. For example, with shape[2]=13 and patch_size_t=4,
this gives ncopy=1, resulting in 14 frames which is still not
divisible by 4, causing the assertion to fail.

The correct formula is `(patch_size_t - latent.shape[2] % patch_size_t) % patch_size_t`
which computes how many frames must be prepended to reach the next
multiple of patch_size_t. The outer modulo handles the already-aligned
case (returns 0 instead of patch_size_t).

Fixes #782
This commit is contained in:
Mr-Neutr0n 2026-02-11 19:35:55 +05:30
parent 7a1af71545
commit 7bc152ff35
2 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
finetune/models/cogvideox_i2v/lora_trainer.py
View File

@ -115,7 +115,7 @@ class CogVideoXI2VLoraTrainer(Trainer):
patch_size_t = self.state.transformer_config.patch_size_t
if patch_size_t is not None:
ncopy = latent.shape[2] % patch_size_t
ncopy = (patch_size_t - latent.shape[2] % patch_size_t) % patch_size_t
# Copy the first frame ncopy times to match patch_size_t
first_frame = latent[:, :, :1, :, :] # Get first frame [B, C, 1, H, W]
latent = torch.cat([first_frame.repeat(1, 1, ncopy, 1, 1), latent], dim=2)

2
finetune/models/cogvideox_t2v/lora_trainer.py
View File

@ -109,7 +109,7 @@ class CogVideoXT2VLoraTrainer(Trainer):
patch_size_t = self.state.transformer_config.patch_size_t
if patch_size_t is not None:
ncopy = latent.shape[2] % patch_size_t
ncopy = (patch_size_t - latent.shape[2] % patch_size_t) % patch_size_t
# Copy the first frame ncopy times to match patch_size_t
first_frame = latent[:, :, :1, :, :] # Get first frame [B, C, 1, H, W]
latent = torch.cat([first_frame.repeat(1, 1, ncopy, 1, 1), latent], dim=2)