Skip to content

[Model]: MiniMaxAI/MiniMax-M3 #1984

Description

@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

Metadata

Metadata

Type

No type

Projects

No projects

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions