[Bug]: All experts of `MiniMaxAI/MiniMax‑M2.1` are uncalibrated

[yiliu30]

Problem Description

This issue does not occur on Qwen/Qwen3‑30B‑A3B; it may be caused by the experts’ linear layer names or FP8 model only?

The get_act_max_hook was missed for dequantized model.
Not sure if it was supported before or not, looks like a feature gap. cc @xin3he @wenhuach21

Reproduction Steps

import torch
from auto_round import AutoRound
model_name = "/storage/yiliu7/MiniMaxAI/MiniMax-M2.1/"
scheme = "FP8_STATIC"
autoround = AutoRound(
            model_name,
            scheme=scheme,
            iters=0,
            low_gpu_mem_usage=True,
            disable_opt_rtn=True,
)
autoround.quantize_and_save(format="auto_round", output_dir="tmp_autoround_qwen3_fix")

Environment Information

model: /storage/yiliu7/MiniMaxAI/MiniMax-M2.1/
commit: 28839a9

Error Logs

- log

model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.86.w2.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.86.w3.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.87.w1.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.87.w2.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.87.w3.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.88.w1.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.88.w2.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.88.w3.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.89.w1.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.89.w2.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.89.w3.input_scale,torch.bfloat16,[1],1.0,1.0
model-00002-of-00007.safetensors,model.layers.1.block_sparse_moe.experts.9.w1.input_scale,torch.bfloat16,[1],1.0,1.0

...

Additional Context

No response