easonysliu e5f6c1ff68 fix: reshape dora_scale before broadcasting in weight_decompose ... In weight_decompose(), the 1D dora_scale tensor [N] divided by the multi-dimensional weight_norm [N, 1, ...] would incorrectly broadcast to [N, N, ...] (outer-product shape) instead of element-wise [N, 1, ...]. This caused shape mismatches when applying DoRA to non-square weight matrices (e.g. MLP layers where d_ff != d_model), while silently producing correct results for square weights (most attention Q/K/V/O). Fix: explicitly reshape dora_scale to match weight_norm's dimensionality before the division. Fixes #12938 Co-Authored-By: Claude (claude-opus-4-6) <noreply@anthropic.com>		2026-03-17 11:17:57 +08:00
..
__init__.py	[Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958)	2026-01-24 22:56:22 -05:00
base.py	fix: reshape dora_scale before broadcasting in weight_decompose	2026-03-17 11:17:57 +08:00
boft.py	[Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958)	2026-01-24 22:56:22 -05:00
bypass.py	[Trainer] training with proper offloading (#12189)	2026-02-10 21:45:19 -05:00
glora.py	[Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958)	2026-01-24 22:56:22 -05:00
loha.py	[Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958)	2026-01-24 22:56:22 -05:00
lokr.py	[Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958)	2026-01-24 22:56:22 -05:00
lora.py	MPDynamic: force load flux img_in weight (Fixes flux1 canny+depth lora crash) (#12446)	2026-02-15 20:30:09 -05:00
oft.py	[Weight-adapter/Trainer] Bypass forward mode in Weight adapter system (#11958)	2026-01-24 22:56:22 -05:00