Kimi-K2-Thinking-MXFP4-AttnFP8

Model Overview

• Model Architecture: Kimi-K2-Thinking
  • Input: Text
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.0
• Transformers: 4.57.6
• Operating System(s): Linux
• Inference Engine: vLLM
• Model Optimizer: AMD-Quark (V0.11.2)
  • Quantized layers: experts, shared_experts, self_attn
  • Weight quantization: MoE OCP MXFP4, Static; self_attn Perchannel, FP8E4M3, Static
  • Activation quantization: MoE OCP MXFP4, Dynamic; self_attn Pertoken, FP8E4M3, Dynamic
• Calibration Dataset: Pile

This model was built with Kimi-K2-Thinking model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from unsloth/Kimi-K2-Thinking-BF16 using AMD-Quark. Both weights and activations were quantized.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="*mlp.gate *lm_head *mlp.gate_proj *mlp.up_proj *mlp.down_proj"

python quantize_quark.py \
    --model_dir unsloth/Kimi-K2-Thinking-BF16 \
    --quant_scheme mxfp4 \
    --layer_quant_scheme '*self_attn*' ptpc_fp8 \
    --exclude_layers  $exclude_layers \
    --output_dir  amd/Kimi-K2-Thinking-MXFP4-AttnFP8 \
    --file2file_quantization

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend.

Evaluation

The model was evaluated on GSM8K benchmarks.

Accuracy

<table> <tr> <td><strong>Benchmark</strong> </td> <td><strong>Kimi-K2-Thinking </strong> </td> <td><strong>Kimi-K2-Thinking-MXFP4-AttnFP8(this model)</strong> </td> <td><strong>Recovery</strong> </td> </tr> <tr> <td>GSM8K (flexible-extract) </td> <td>94.16 </td> <td>92.95 </td> <td>98.71% </td> </tr> </table>

Reproduction

The GSM8K results were obtained using the lm-evaluation-harness framework, based on the Docker image rocm/vllm-private:vllm_dev_base_mxfp4_20260122, with vLLM, lm-eval and amd-quark compiled and installed from source inside the image.

Launching server

export VLLM_ATTENTION_BACKEND="TRITON_MLA"
export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=0

vllm serve amd/Kimi-K2-Thinking-MXFP4-AttnFP8 \
  --tensor-parallel-size 8 \
  --enable-auto-tool-choice \
  --tool-call-parser kimi_k2 \
  --reasoning-parser kimi_k2 \
  --trust-remote-code

Evaluating model in a new terminal

lm_eval \
  --model local-completions \
  --model_args "model=amd/Kimi-K2-Thinking-MXFP4-AttnFP8,base_url=http://0.0.0.0:8000/v1/completions,tokenized_requests=False,tokenizer_backend=None,num_concurrent=32" \
  --tasks gsm8k \
  --num_fewshot 5 \
  --batch_size 1

License

Modifications Copyright(c) 2025 Advanced Micro Devices, Inc. All rights reserved.