Propose to refactor `norm_out` in `SkyReelsV2Transformer3DModel` for device handling in multi-GPU case#11850

Open

tolgacangoz wants to merge 31 commits intohuggingface:mainfrom

tolgacangoz:transfer-shift_scale_norm-to-AdaLayerNorm

Commits on Jul 3, 2025

Commits on Jul 10, 2025

Commits on Jul 18, 2025

Merge branch 'main' into transfer-shift_scale_norm-to-AdaLayerNorm
tolgacangoz
authored

Commits on Aug 16, 2025

Commits on Aug 17, 2025

revert
tolgacangoz
committed
Implement `SkyReelsV2AdaLayerNorm` for timestep embedding modulation and normalization
tolgacangoz
committed
style
tolgacangoz
committed
Enhances backward compatibility by converting deprecated `scale_shift_table` to `SkyReelsV2AdaLayerNorm` format in `_load_from_state_dict`
tolgacangoz
committed
Adds missing keys to ignore on load and adjusts identity matrix creation for device and dtype consistency
tolgacangoz
committed
Fix device and dtype handling for identity matrix and scale_shift_table in SkyReelsV2Transformer3DModel
tolgacangoz
committed
Merge branch 'main' into transfer-shift_scale_norm-to-AdaLayerNorm
tolgacangoz
authored
Update normalization.py
tolgacangoz
authored

Commits on Aug 18, 2025

Commits on Aug 19, 2025

Commits on Aug 20, 2025

Merge branch 'main' into transfer-shift_scale_norm-to-AdaLayerNorm
tolgacangoz
authored

Commits on Aug 26, 2025

Merge branch 'main' into transfer-shift_scale_norm-to-AdaLayerNorm
tolgacangoz
authored

Commits on Sep 3, 2025

Merge branch 'main' into transfer-shift_scale_norm-to-AdaLayerNorm
tolgacangoz
authored