diff --git a/unsloth/kernels/fast_lora.py b/unsloth/kernels/fast_lora.py index 4f91254e6c..e76173b479 100644 --- a/unsloth/kernels/fast_lora.py +++ b/unsloth/kernels/fast_lora.py @@ -123,8 +123,8 @@ class LoRA_MLP(torch.autograd.Function): g = g .view(-1, g .shape[-1]) dtype = X.dtype - DW_f = (D @ W.T * f) - DW_dfg = (D @ W.T * df * g) + DW_f = (dY @ W.T * f) + DW_dfg = (dY @ W.T * df * g) se = torch.nn.functional.sigmoid(e) f = e * se h = f * g