File size: 3,061 Bytes
7415d2f
 
 
 
636197a
 
7415d2f
 
636197a
7415d2f
636197a
 
 
 
189a20b
636197a
 
dc8a62e
f8e491f
 
636197a
 
 
 
 
 
dc8a62e
 
 
 
 
 
 
 
 
 
 
 
 
 
636197a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e36ddc2
636197a
 
 
 
 
 
 
 
 
 
 
 
60ea990
189a20b
636197a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
---
license: other
license_name: modified-mit
license_link: LICENSE
base_model:
- moonshotai/Kimi-K2-Thinking
---

# Model Overview

- **Model Architecture:** Kimi-K2-Thinking
  - **Input:** Text
  - **Output:** Text
- **Supported Hardware Microarchitecture:** AMD MI350/MI355
- **ROCm:** 7.0
- **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) (V0.11.1)
  - **Weight quantization:** MOE-only, OCP MXFP4, Static 
  - **Activation quantization:** MOE-only, OCP MXFP4, Dynamic
- **Calibration Dataset:** [Pile](https://huggingface.co/datasets/mit-han-lab/pile-val-backup)

This model was built with Kimi-K2-Thinking model by applying [AMD-Quark](https://quark.docs.amd.com/latest/index.html) for MXFP4 quantization.

# Model Quantization

The model was quantized from [unsloth/Kimi-K2-Thinking-BF16](https://huggingface.co/unsloth/Kimi-K2-Thinking-BF16) using [AMD-Quark](https://quark.docs.amd.com/latest/index.html). The weights and activations are quantized to MXFP4.

**Quantization scripts:**
```
cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="*self_attn* *mlp.gate *lm_head *mlp.gate_proj *mlp.up_proj *mlp.down_proj *shared_experts*"

python quantize_quark.py \
    --model_dir unsloth/Kimi-K2-Thinking-BF16 \
    --quant_scheme mxfp4 \
    --exclude_layers  $exclude_layers \
    --output_dir amd/Kimi-K2-Thinking-MXFP4 \
    --file2file_quantization
```

# Deployment
### Use with vLLM

This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) 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(this model)</strong>
   </td>
   <td><strong>Recovery</strong>
   </td>
  </tr>
  <tr>
   <td>GSM8K (strict-match)
   </td>
   <td>94.16
   </td>
   <td>93.48
   </td>
   <td>99.28%
   </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 \
  --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,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.