Summary

DoRA (use_dora=True) adapters could not be merged through the safetensors path. create_lora_statistics did not capture the lora_magnitude_vector, so the key-consistency check raised Unsloth: Extracted keys = {...lora_magnitude_vector.default.weight} do not match!, and even past that the merge math applied only W0 + alpha*BA, dropping the magnitude.

This adds DoRA support for the dense merge path (attention / MLP / any dense linear, including the language and vision/audio towers of multimodal models):

• LoraStats gains a magnitude field; create_lora_statistics captures lora_magnitude_vector.default.
• assert_same_keys ignores the magnitude tensor (folded into the weight, omitted from the merged checkpoint), like lora_A / lora_B.
• _merge_lora folds it in as (m / ||W0 + alpha*BA||_row) * (W0 + alpha*BA), i.e. one L2 norm per output row over the input dim, matching PEFT's DoRA merge.

MoE-expert DoRA is explicitly refused with a clear error (the per-expert / fused expert helpers fold only the LoRA delta, not the magnitude), so a DoRA MoE merge fails loud instead of silently dropping the magnitude.

Test

tests/test_dora_merge.py:

• dense DoRA merge equals PEFT's own merge_and_unload (atol 1e-4) and actually changes the weight;
• plain LoRA (use_dora=False) is unchanged (magnitude is None);
• DoRA on a MoE expert raises.

Full merge e2e suite (dense + MoE fused/per-expert + resized) stays green (65 passed, 6 skipped).