Error in LoRA training on MACE-POLAR1

[jylim-4723]

Bug in merging LoRA weights into base model (polar-1-m)

line 88, in _extract_weight_blocks
size = instr.path_shape[0] * instr.path_shape[1]
IndexError: tuple index out of range

MACE version: 0.3.16
CUDA version: 12.6
Torch version: 2.13.0