README.md

---
pipeline_tag: image-text-to-text
license: other
license_name: minimax-community
license_link: LICENSE
library_name: transformers
tags:
- multimodal
- moe
- agent
- coding
- video
base_model:
- MiniMaxAI/MiniMax-M3
---

# Model Overview

- **Model Architecture:** MiniMaxM3SparseForConditionalGeneration
  - **Input:** Text, Image
  - **Output:** Text
- **Supported Hardware Microarchitecture:** AMD MI350/MI355
- **ROCm**: 7.1.1
- **PyTorch**: 2.10.0
- **Transformers**: 5.2.0
- **Operating System(s):** Linux
- **Inference Engine:** [vLLM](https://docs.vllm.ai/en/latest/)
- **Model Optimizer:** [AMD-Quark](https://quark.docs.amd.com/latest/index.html)
  - **Weight quantization:** OCP MXFP4, Static
  - **Activation quantization:** OCP MXFP4, Dynamic


# Model Quantization

The model was quantized from [MiniMaxAI/MiniMax-M3](https://huggingface.co/MiniMaxAI/MiniMax-M3) using [AMD-Quark](https://quark.docs.amd.com/latest/index.html). The weights are quantized to MXFP4 and activations are quantized to MXFP4.


**Quantization scripts:**
```python
from quark.torch import LLMTemplate, ModelQuantizer

# --- Register template ---
minimax_m3_vl_template = LLMTemplate(
    model_type="minimax_m3_vl",
    kv_layers_name=["*language_model.*k_proj", "*language_model.*v_proj"],
    q_layer_name="*language_model.*q_proj",
    exclude_layers_name=[
        "*lm_head",
        "*vision_tower*",
        "*multi_modal_projector*",
        "*patch_merge_mlp*",
        "*block_sparse_moe.gate",
        "*self_attn*",
    ],
)
LLMTemplate.register_template(minimax_m3_vl_template)
print(f"[INFO]: Registered template '{minimax_m3_vl_template.model_type}'")

# --- Configuration ---
model_dir = "MiniMaxAI/MiniMax-M3"
output_dir = "amd/MiniMax-M3-MXFP4"
quant_scheme = "mxfp4"
exclude_layers = [
    "*lm_head",
    "*vision_tower*",
    "*multi_modal_projector*",
    "*patch_merge_mlp*",
    "*block_sparse_moe.gate",
    "*self_attn*",
    "*mlp.gate_proj",
    "*mlp.up_proj",
    "*mlp.down_proj",
]

# --- Build quant config from template ---
template = LLMTemplate.get("minimax_m3_vl")
quant_config = template.get_config(scheme=quant_scheme, exclude_layers=exclude_layers)

# --- File-to-file quantization (memory-efficient, no full model loading) ---
quantizer = ModelQuantizer(quant_config)
quantizer.direct_quantize_checkpoint(
    pretrained_model_path=model_dir,
    save_path=output_dir,
)
print(f"[INFO]: Quantization complete. Output saved to {output_dir}")
```

# Evaluation
The model was evaluated on gsm8k benchmarks using the vllm framework.

### Accuracy

<table>
  <tr>
   <td><strong>Benchmark</strong>
   </td>
   <td><strong>MiniMaxAI/MiniMax-M3 </strong>
   </td>
   <td><strong>amd/MiniMax-M3-MXFP4(this model)</strong>
   </td>
   <td><strong>Recovery</strong>
   </td>
  </tr>
  <tr>
   <td>gsm8k (flexible-extract) 
   </td>
   <td>95.30
   </td>
   <td>94.19
   </td>
   <td>98.84%
   </td>
  </tr>
</table>


### Reproduction

The GSM8K results were obtained using the lm-eval framework, based on the
Docker image `rocm/pytorch-private:vllm-hy-mm-06112026`. The vLLM shipped in
that image was used as-is, with the patch from this PR ([#45794](https://github.com/vllm-project/vllm/pull/45794/changes)) applied on top.

#### Launching server
```bash
vllm serve /mnt/amd/MiniMax-M3-MXFP4 \
  --trust-remote-code \
  --block-size 128 \
  --tensor-parallel-size 8 \
  --attention-backend TRITON_ATTN \
  --mm-encoder-tp-mode data \
  --mm-encoder-attn-backend ROCM_AITER_FA \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice \
  --reasoning-parser minimax_m3 \
  --moe-backend emulation
```


#### Evaluating model in a new terminal
```bash
lm_eval \
  --model local-chat-completions \
  --model_args "model=/mnt/amd/MiniMax-M3-MXFP4,base_url=http://127.0.0.1:8000/v1/chat/completions,num_concurrent=32,max_gen_toks=16384" \
  --tasks gsm8k \
  --num_fewshot 5 \
  --batch_size 1 \
  --apply_chat_template \
  --fewshot_as_multiturn
```