[Model]: MiniMaxAI/MiniMax-M3

[lkk12014402]

Model name

MiniMaxAI/MiniMax-M3

Model variant / size

427B

Model link

https://huggingface.co/MiniMaxAI/MiniMax-M3

What do you want AutoRound to support?

Quantization / rounding workflow support for this model

Motivation & use case

quantize model with non model-free method

The config.json generated by the quantized model appears to be incorrect. Could you explain why the following configuration is present?

"layer_types": [
    "full_attention",
    "full_attention",
    "full_attention",
    "minimax_m3_sparse",
    "minimax_m3_sparse",
    "minimax_m3_sparse",
    "minimax_m3_sparse"
]

This causes VLLM to fail when loading the model. However, if I manually remove the layer_types field (or the related entries), the model can be loaded successfully, but the GSM8K accuracy drops to 0.

Is this a known issue in the quantization/export process, or is there a problem with how config.json is generated?

Environment (optional but helpful)

No response

Current blockers / errors (if any)

No response

Definition of Done (acceptance criteria)

No response

Additional context

No response

[lkk12014402]

https://huggingface.co/INCModelSharing/MiniMax-M3-MXFP4-Mixed-CT-AutoRound-Tuning

[lkk12014402]

MXFP8

IFBench, num_repeats: 8 temperature=0.6, max_seq=8192

[repeat 1/8]
Generating (seed=42): 54%|████████████████████████████████████▏ | 162/300 [20:37<44:43, 19.44s/it]

Generating (seed=42): 100%|███████████████████████████████████████████████████████████████████| 300/300 [53:55<00:00, 10.79s/it]
strict prompt-level=0.5167 instruction-level=0.5087
loose prompt-level=0.5767 instruction-level=0.5669

[repeat 2/8]
Generating (seed=43): 100%|███████████████████████████████████████████████████████████████████| 300/300 [53:20<00:00, 10.67s/it]
strict prompt-level=0.4800 instruction-level=0.4767
loose prompt-level=0.5367 instruction-level=0.5262

[repeat 3/8]
Generating (seed=44): 100%|███████████████████████████████████████████████████████████████████| 300/300 [52:07<00:00, 10.42s/it]
strict prompt-level=0.4967 instruction-level=0.5029
loose prompt-level=0.5600 instruction-level=0.5581

[repeat 4/8]
Generating (seed=45): 100%|███████████████████████████████████████████████████████████████████| 300/300 [52:10<00:00, 10.44s/it]
strict prompt-level=0.4900 instruction-level=0.4913
loose prompt-level=0.5467 instruction-level=0.5436

[repeat 5/8]
Generating (seed=46): 100%|███████████████████████████████████████████████████████████████████| 300/300 [54:35<00:00, 10.92s/it]
strict prompt-level=0.5033 instruction-level=0.5058
loose prompt-level=0.5567 instruction-level=0.5552

[repeat 6/8]
Generating (seed=47): 100%|███████████████████████████████████████████████████████████████████| 300/300 [53:10<00:00, 10.64s/it]
strict prompt-level=0.4967 instruction-level=0.4913
loose prompt-level=0.5400 instruction-level=0.5291