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
• Model Optimizer: AMD-Quark
  • Weight quantization: OCP MXFP4, Static; self_attn Perchannel, FP8E4M3, Static
  • Activation quantization: OCP MXFP4, Dynamic; self_attn Pertoken, FP8E4M3, Dynamic

Model Quantization

The model was quantized from MiniMaxAI/MiniMax-M3 using AMD-Quark. The weights are quantized to MXFP4 and activations are quantized to MXFP4, and self_attn layers are quantized to PTPC-FP8.

Quantization scripts:

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",
    ],
)
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-AttnFP8"
quant_scheme = "mxfp4"
# Per-layer override: self_attn (q/k/v/o_proj) -> ptpc_fp8 instead of mxfp4.
# Equivalent to:  --layer_quant_scheme '*self_attn*' ptpc_fp8
layer_config = {
    "*self_attn*": "ptpc_fp8",
}
exclude_layers = [
    "*lm_head",
    "*vision_tower*",
    "*multi_modal_projector*",
    "*patch_merge_mlp*",
    "*block_sparse_moe.gate",
    "*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,
    layer_config=layer_config,
    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

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) applied on top.

Before running the evaluation, install the required packages:

pip install transformers==5.12.0
pip install "lm-eval[api]"

Launching server

vllm serve /mnt/amd/MiniMax-M3-MXFP4-AttnFP8 \
  --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

lm_eval \
  --model local-chat-completions \
  --model_args "model=/mnt/amd/MiniMax-M3-MXFP4-AttnFP8,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