Model Architecture: Kimi-K2.5
- Input: Text
- Output: Text
Supported Hardware Microarchitecture: AMD MI300/MI325/MI350/MI355
ROCm: 7.1.0
Operating System(s): Linux
Inference Engine: vLLM
Model Optimizer: AMD-Quark
- Weight quantization: MOE-only, INT4 Per-Channel & FP8E4M3, Static
- Activation quantization: MOE-only, FP8E4M3, Dynamic

This model was built with Kimi-K2.5 model by applying AMD-Quark for INT4-FP8 quantization.

Model Quantization

The model was quantized from moonshotai/Kimi-K2.5 using AMD-Quark. The weights and activations are quantized to INT4-FP8.

python

#!/usr/bin/env python3
"""Kimi-K2.5 W4A8 re-quantization via AMD Quark (File-to-File Quantization)."""

import argparse
import os
from quark.torch.quantization.config.config import (
    FP8E4M3PerTensorSpec, Int4PerChannelSpec, ProgressiveSpec,
    QConfig, QLayerConfig,
)
from quark.torch.quantization.api import ModelQuantizer


def get_config():
    exclude_layers = [
        "*self_attn*", "*mlp.gate", "*lm_head",
        "*mlp.gate_proj", "*mlp.up_proj", "*mlp.down_proj",
        "*shared_experts*", "*mm_projector*", "*vision_tower*",
    ]

    input_spec = FP8E4M3PerTensorSpec(
        observer_method="min_max", scale_type="float32", is_dynamic=True,
    ).to_quantization_spec()

    weight_spec = ProgressiveSpec(
        first_stage=FP8E4M3PerTensorSpec(
            observer_method="min_max", scale_type="float32", is_dynamic=False,
        ),
        second_stage=Int4PerChannelSpec(
            symmetric=True, scale_type="float32",
            round_method="half_even", is_dynamic=False, ch_axis=0,
        ),
    ).to_quantization_spec()

    return QConfig(
        global_quant_config=QLayerConfig(input_tensors=input_spec, weight=weight_spec),
        exclude=exclude_layers,
    )


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--model-path", type=str, required=True,
                        help="Path to moonshotai/Kimi-K2.5 checkpoint")
    parser.add_argument("--export-path", type=str, required=True,
                        help="Path to save quantized output")
    args = parser.parse_args()

    quantizer = ModelQuantizer(get_config())
    quantizer.direct_quantize_checkpoint(
        pretrained_model_path=args.model_path,
        save_path=args.export_path,
    )
    print("[INFO] Quantization completed")

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend.

Evaluation

The model was evaluated on GSM8K benchmarks.

Accuracy

Benchmark	Kimi-K2.5	Kimi-K2.5-W4A8(this model)	Recovery
GSM8K (flexible-extract)	94.09	93.40	99.27%

Reproduction

The GSM8K results were obtained using the lm-evaluation-harness framework, based on the Docker image vllm/vllm-openai-rocm:v0.14.0.

Install the vLLM (commit ecb4f822091a64b5084b3a4aff326906487a363f) and lm-eval (Version: 0.4.10) in container first.

git clone https://github.com/vllm-project/vllm.git
cd vllm
python3 setup.py develop

pip install lm-eval

Launching server

VLLM_ROCM_USE_AITER_MLA=0 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=0 VLLM_ROCM_USE_AITER_FP4BMM=0 vllm serve amd/Kimi-K2.5-W4A8 \
  --tensor-parallel-size 8 \
  --mm-encoder-tp-mode data \
  --tool-call-parser kimi_k2 \
  --reasoning-parser kimi_k2 \
  --trust-remote-code \
  --enforce-eager

Evaluating model in a new terminal

lm_eval \
  --model local-completions \
  --model_args "model=amd/Kimi-K2.5-W4A8,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

amd/Kimi-K2.5-W4A8

Model Overview