Modelpublic

amd/Kimi-K2.5-MXFP4-AttnFP8

sourceHugging Faceotherupdated 5mo agoView on Hugging Face

1likes103downloads

Model Card

Model Overview

—Model Architecture: Kimi-K2.5
—Input: Text
—Output: Text
—Supported Hardware Microarchitecture: AMD MI350/MI355
—ROCm: 7.1.0
—Transformers: 4.57.6
—Operating System(s): Linux
—Inference Engine: vLLM
—Model Optimizer: AMD-Quark (V0.11.2)
—Quantized layers: layers.0.mlp, experts, shared_experts, self_attn
—Weight quantization: OCP MXFP4, Static; self_attn Perchannel, FP8E4M3, Static
—Activation quantization: OCP MXFP4, Dynamic; self_attn Pertoken, FP8E4M3, Dynamic
—Calibration Dataset: Pile

This model was built with Kimi-K2.5 model by applying AMD-Quark for MXFP4 quantization and PTPC-FP8 quantization.

Model Quantization

The model was quantized from moonshotai/Kimi-K2.5 using AMD-Quark. The weights and activations are quantized to MXFP4, and self_attn layers are quantized to PTPC-FP8.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="*mlp.gate *lm_head *mm_projector* *vision_tower*"

python3 quantize_quark.py \
  --model_dir moonshotai/Kimi-K2.5 \
  --quant_scheme mxfp4 \
  --layer_quant_scheme '*self_attn*' ptpc_fp8 \
  --exclude_layers $exclude_layers \
  --output_dir amd/Kimi-K2.5-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-K2.5 </strong> </td> <td><strong>Kimi-K2.5-MXFP4-AttnFP8(this model) </strong> </td> <td><strong>Recovery</strong> </td> </tr> <tr> <td>GSM8K (flexible-extract) </td> <td>94.09 </td> <td>93.56 </td> <td>99.44% </td> </tr> </table>

Reproduction

The GSM8K results were obtained using the lm-evaluation-harness framework, based on the Docker image vllm/vllm-openai-rocm:v0.17.0.

Install the lm-eval (Version: 0.4.11) in container first.

pip install lm-eval
pip install lm-eval[api]

Launching server

export VLLM_ROCM_USE_AITER=1

vllm serve amd/Kimi-K2.5-MXFP4-AttnFP8 -tp 4 \
  --mm-encoder-tp-mode data \
  --tool-call-parser kimi_k2 \
  --reasoning-parser kimi_k2 \
  --enforce-eager \
  --trust-remote-code

Evaluating model in a new terminal

lm_eval \
  --model local-completions \
  --model_args "model=amd/Kimi-K2.5-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