[Bug]: MTP logic is unaware of Transformers' layer renaming and quantizes the entire model again

[wenhuach21]

Problem Description

auto-round ./ --eval --eval_bs 16 --tasks lambada_openai

1 unexpected keys

]
[transformers] NemotronHForCausalLM LOAD REPORT from: ./
Key | Status | Details
-----------------------------------------+------------+--------
model.layers.{1...51}.mixer.gate.qzeros | UNEXPECTED |
model.layers.{1...51}.mixer.gate.qweight | UNEXPECTED |
model.layers.{1...51}.mixer.gate.scales | UNEXPECTED |

2 gate_proj issue

File "/home/wenhuach/miniforge3/envs/autoround/lib/python3.13/site-packages/auto_round/modeling/fused_moe/moe_experts_interface.py", line 302, in linear_loop_experts_forward
final_hidden_states = _run_experts_with_routes(top_k_index, top_k_weights)
File "/home/wenhuach/miniforge3/envs/autoround/lib/python3.13/site-packages/auto_round/modeling/fused_moe/moe_experts_interface.py", line 279, in _run_experts_with_routes
gate_out = expert.gate_proj(expert_input) # (num_samples, intermediate_dim)
^^^^^^^^^^^^^^^^
File "/home/wenhuach/miniforge3/envs/autoround/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1972, in getattr
raise AttributeError(
f"'{type(self).name}' object has no attribute '{name}'"
)
AttributeError: '_ExpertContainer' object has no attribute 'gate_proj'

Reproduction Steps

~

Environment Information

No response

Error Logs

Additional Context

No response

[wenhuach21]

1 belongs to XIN. The checkpoints are under backbone.layers. Transformers renames them to model.layers. The MTP logic quantizes all checkpoints using the original naming scheme. After the evaluation phase, Transformers renames backbone.layers again. I think the rtn based quantized will overwrite the tuned version. @xin3he

[wenhuach21]

issue 2 fix is here fix_Nemotron_gate