Repository: MoonshotAI/Kimi-K2
Branch: main
Commit: 1b4022bbb718
Files: 4
Total size: 47.0 KB
Directory structure:
gitextract_zye998ae/
├── LICENSE
├── README.md
└── docs/
├── deploy_guidance.md
└── tool_call_guidance.md
================================================
FILE CONTENTS
================================================
================================================
FILE: LICENSE
================================================
Modified MIT License
Copyright (c) 2025 Moonshot AI
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the “Software”), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
Our only modification part is that, if the Software (or any derivative works
thereof) is used for any of your commercial products or services that have
more than 100 million monthly active users, or more than 20 million US dollars
(or equivalent in other currencies) in monthly revenue, you shall prominently
display "Kimi K2" on the user interface of such product or service.
================================================
FILE: README.md
================================================
📰 Tech Blog | 📄 Full Report
## 1. Model Introduction
Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters. Trained with the Muon optimizer, Kimi K2 achieves exceptional performance across frontier knowledge, reasoning, and coding tasks while being meticulously optimized for agentic capabilities.
### Key Features
- Large-Scale Training: Pre-trained a 1T parameter MoE model on 15.5T tokens with zero training instability.
- MuonClip Optimizer: We apply the Muon optimizer to an unprecedented scale, and develop novel optimization techniques to resolve instabilities while scaling up.
- Agentic Intelligence: Specifically designed for tool use, reasoning, and autonomous problem-solving.
### Model Variants
- **Kimi-K2-Base**: The foundation model, a strong start for researchers and builders who want full control for fine-tuning and custom solutions.
- **Kimi-K2-Instruct**: The post-trained model, best for drop-in, general-purpose chat and agentic experiences. It is a reflex-grade model without long thinking.
## 2. Model Summary
| | |
|:---:|:---:|
| **Architecture** | Mixture-of-Experts (MoE) |
| **Total Parameters** | 1T |
| **Activated Parameters** | 32B |
| **Number of Layers** (Dense layer included) | 61 |
| **Number of Dense Layers** | 1 |
| **Attention Hidden Dimension** | 7168 |
| **MoE Hidden Dimension** (per Expert) | 2048 |
| **Number of Attention Heads** | 64 |
| **Number of Experts** | 384 |
| **Selected Experts per Token** | 8 |
| **Number of Shared Experts** | 1 |
| **Vocabulary Size** | 160K |
| **Context Length** | 128K |
| **Attention Mechanism** | MLA |
| **Activation Function** | SwiGLU |
## 3. Evaluation Results
#### Instruction model evaluation results
| Benchmark |
Metric |
Kimi K2 Instruct |
DeepSeek-V3-0324 |
Qwen3-235B-A22B
(non-thinking) |
Claude Sonnet 4
(w/o extended thinking) |
Claude Opus 4
(w/o extended thinking) |
GPT-4.1 |
Gemini 2.5 Flash
Preview (05-20) |
| Coding Tasks |
LiveCodeBench v6
(Aug 24 - May 25) |
Pass@1 |
53.7 |
46.9 |
37.0 |
48.5 |
47.4 |
44.7 |
44.7 |
| OJBench |
Pass@1 |
27.1 |
24.0 |
11.3 |
15.3 |
19.6 |
19.5 |
19.5 |
| MultiPL-E |
Pass@1 |
85.7 |
83.1 |
78.2 |
88.6 |
89.6 |
86.7 |
85.6 |
SWE-bench Verified
(Agentless Coding) |
Single Patch w/o Test (Acc) |
51.8 |
36.6 |
39.4 |
50.2 |
53.0 |
40.8 |
32.6 |
SWE-bench Verified
(Agentic Coding) |
Single Attempt (Acc) |
65.8 |
38.8 |
34.4 |
72.7* |
72.5* |
54.6 |
— |
| Multiple Attempts (Acc) |
71.6 |
— |
— |
80.2 |
79.4* |
— |
— |
SWE-bench Multilingual
(Agentic Coding) |
Single Attempt (Acc) |
47.3 |
25.8 |
20.9 |
51.0 |
— |
31.5 |
— |
| TerminalBench |
Inhouse Framework (Acc) |
30.0 |
— |
— |
35.5 |
43.2 |
8.3 |
— |
| Terminus (Acc) |
25.0 |
16.3 |
6.6 |
— |
— |
30.3 |
16.8 |
| Aider-Polyglot |
Acc |
60.0 |
55.1 |
61.8 |
56.4 |
70.7 |
52.4 |
44.0 |
| Tool Use Tasks |
| Tau2 retail |
Avg@4 |
70.6 |
69.1 |
57.0 |
75.0 |
81.8 |
74.8 |
64.3 |
| Tau2 airline |
Avg@4 |
56.5 |
39.0 |
26.5 |
55.5 |
60.0 |
54.5 |
42.5 |
| Tau2 telecom |
Avg@4 |
65.8 |
32.5 |
22.1 |
45.2 |
57.0 |
38.6 |
16.9 |
| AceBench |
Acc |
76.5 |
72.7 |
70.5 |
76.2 |
75.6 |
80.1 |
74.5 |
| Math & STEM Tasks |
| AIME 2024 |
Avg@64 |
69.6 |
59.4* |
40.1* |
43.4 |
48.2 |
46.5 |
61.3 |
| AIME 2025 |
Avg@64 |
49.5 |
46.7 |
24.7* |
33.1* |
33.9* |
37.0 |
46.6 |
| MATH-500 |
Acc |
97.4 |
94.0* |
91.2* |
94.0 |
94.4 |
92.4 |
95.4 |
| HMMT 2025 |
Avg@32 |
38.8 |
27.5 |
11.9 |
15.9 |
15.9 |
19.4 |
34.7 |
| CNMO 2024 |
Avg@16 |
74.3 |
74.7 |
48.6 |
60.4 |
57.6 |
56.6 |
75.0 |
| PolyMath-en |
Avg@4 |
65.1 |
59.5 |
51.9 |
52.8 |
49.8 |
54.0 |
49.9 |
| ZebraLogic |
Acc |
89.0 |
84.0 |
37.7* |
73.7 |
59.3 |
58.5 |
57.9 |
| AutoLogi |
Acc |
89.5 |
88.9 |
83.3 |
89.8 |
86.1 |
88.2 |
84.1 |
| GPQA-Diamond |
Avg@8 |
75.1 |
68.4* |
62.9* |
70.0* |
74.9* |
66.3 |
68.2 |
| SuperGPQA |
Acc |
57.2 |
53.7 |
50.2 |
55.7 |
56.5 |
50.8 |
49.6 |
Humanity's Last Exam
(Text Only) |
- |
4.7 |
5.2 |
5.7 |
5.8 |
7.1 |
3.7 |
5.6 |
| General Tasks |
| MMLU |
EM |
89.5 |
89.4 |
87.0 |
91.5 |
92.9 |
90.4 |
90.1 |
| MMLU-Redux |
EM |
92.7 |
90.5 |
89.2 |
93.6 |
94.2 |
92.4 |
90.6 |
| MMLU-Pro |
EM |
81.1 |
81.2* |
77.3 |
83.7 |
86.6 |
81.8 |
79.4 |
| IFEval |
Prompt Strict |
89.8 |
81.1 |
83.2* |
87.6 |
87.4 |
88.0 |
84.3 |
| Multi-Challenge |
Acc |
54.1 |
31.4 |
34.0 |
46.8 |
49.0 |
36.4 |
39.5 |
| SimpleQA |
Correct |
31.0 |
27.7 |
13.2 |
15.9 |
22.8 |
42.3 |
23.3 |
| Livebench |
Pass@1 |
76.4 |
72.4 |
67.6 |
74.8 |
74.6 |
69.8 |
67.8 |
• Bold denotes global SOTA, and underlined denotes open-source SOTA.
• Data points marked with * are directly from the model's tech report or blog.
• All metrics, except for SWE-bench Verified (Agentless), are evaluated with an 8k output token length. SWE-bench Verified (Agentless) is limited to a 16k output token length.
• Kimi K2 achieves 65.8% pass@1 on the SWE-bench Verified tests with bash/editor tools (single-attempt patches, no test-time compute). It also achieves a 47.3% pass@1 on the SWE-bench Multilingual tests under the same conditions. Additionally, we report results on SWE-bench Verified tests (71.6%) that leverage parallel test-time compute by sampling multiple sequences and selecting the single best via an internal scoring model.
• To ensure the stability of the evaluation, we employed avg@k on the AIME, HMMT, CNMO, PolyMath-en, GPQA-Diamond, EvalPlus, Tau2.
• Some data points have been omitted due to prohibitively expensive evaluation costs.
---
#### Base model evaluation results
| Benchmark |
Metric |
Shot |
Kimi K2 Base |
Deepseek-V3-Base |
Qwen2.5-72B |
Llama 4 Maverick |
| General Tasks |
| MMLU |
EM |
5-shot |
87.8 |
87.1 |
86.1 |
84.9 |
| MMLU-pro |
EM |
5-shot |
69.2 |
60.6 |
62.8 |
63.5 |
| MMLU-redux-2.0 |
EM |
5-shot |
90.2 |
89.5 |
87.8 |
88.2 |
| SimpleQA |
Correct |
5-shot |
35.3 |
26.5 |
10.3 |
23.7 |
| TriviaQA |
EM |
5-shot |
85.1 |
84.1 |
76.0 |
79.3 |
| GPQA-Diamond |
Avg@8 |
5-shot |
48.1 |
50.5 |
40.8 |
49.4 |
| SuperGPQA |
EM |
5-shot |
44.7 |
39.2 |
34.2 |
38.8 |
| Coding Tasks |
| LiveCodeBench v6 |
Pass@1 |
1-shot |
26.3 |
22.9 |
21.1 |
25.1 |
| EvalPlus |
Pass@1 |
- |
80.3 |
65.6 |
66.0 |
65.5 |
| Mathematics Tasks |
| MATH |
EM |
4-shot |
70.2 |
60.1 |
61.0 |
63.0 |
| GSM8k |
EM |
8-shot |
92.1 |
91.7 |
90.4 |
86.3 |
| Chinese Tasks |
| C-Eval |
EM |
5-shot |
92.5 |
90.0 |
90.9 |
80.9 |
| CSimpleQA |
Correct |
5-shot |
77.6 |
72.1 |
50.5 |
53.5 |
• We only evaluate open-source pretrained models in this work. We report results for Qwen2.5-72B because the base checkpoint for Qwen3-235B-A22B was not open-sourced at the time of our study.
• All models are evaluated using the same evaluation protocol.
## 4. Deployment
> [!Note]
> You can access Kimi K2's API on https://platform.moonshot.ai , we provide an OpenAI/Anthropic-compatible API for you.
>
> The Anthropic-compatible API maps temperature by `real_temperature = request_temperature * 0.6` for better compatiblity with existing applications.
Our model checkpoints are stored in block-fp8 format, you can find it on [Huggingface](https://huggingface.co/moonshotai/Kimi-K2-Instruct).
Currently, it is recommended to run Kimi-K2 on the following inference engines:
* vLLM
* SGLang
* KTransformers
* TensorRT-LLM
Deployment examples for vLLM and SGLang can be found in the [Model Deployment Guide](docs/deploy_guidance.md).
---
## 5. Model Usage
### Chat Completion
Once the local inference service is set up, you can interact with it through the chat endpoint:
```python
def simple_chat(client: OpenAI, model_name: str):
messages = [
{"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
{"role": "user", "content": [{"type": "text", "text": "Please give a brief self-introduction."}]},
]
response = client.chat.completions.create(
model=model_name,
messages=messages,
stream=False,
temperature=0.6,
max_tokens=256
)
print(response.choices[0].message.content)
```
> [!NOTE]
> The recommended temperature for Kimi-K2-Instruct is `temperature = 0.6`.
> If no special instructions are required, the system prompt is a good default.
---
### Tool Calling
Kimi-K2-Instruct has strong tool-calling capabilities.
To enable them, you need to pass the list of available tools in each request, then the model will autonomously decide when and how to invoke them.
The following example demonstrates calling a weather tool end-to-end:
```python
# Your tool implementation
def get_weather(city: str) -> dict:
return {"weather": "Sunny"}
# Tool schema definition
tools = [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Retrieve current weather information. Call this when the user asks about the weather.",
"parameters": {
"type": "object",
"required": ["city"],
"properties": {
"city": {
"type": "string",
"description": "Name of the city"
}
}
}
}
}]
# Map tool names to their implementations
tool_map = {
"get_weather": get_weather
}
def tool_call_with_client(client: OpenAI, model_name: str):
messages = [
{"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
{"role": "user", "content": "What's the weather like in Beijing today? Use the tool to check."}
]
finish_reason = None
while finish_reason is None or finish_reason == "tool_calls":
completion = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.6,
tools=tools, # tool list defined above
tool_choice="auto"
)
choice = completion.choices[0]
finish_reason = choice.finish_reason
if finish_reason == "tool_calls":
messages.append(choice.message)
for tool_call in choice.message.tool_calls:
tool_call_name = tool_call.function.name
tool_call_arguments = json.loads(tool_call.function.arguments)
tool_function = tool_map[tool_call_name]
tool_result = tool_function(**tool_call_arguments)
print("tool_result:", tool_result)
messages.append({
"role": "tool",
"tool_call_id": tool_call.id,
"name": tool_call_name,
"content": json.dumps(tool_result)
})
print("-" * 100)
print(choice.message.content)
```
The `tool_call_with_client` function implements the pipeline from user query to tool execution.
This pipeline requires the inference engine to support Kimi-K2’s native tool-parsing logic.
For streaming output and manual tool-parsing, see the [Tool Calling Guide](docs/tool_call_guidance.md).
---
## 6. License
Both the code and the model weights are released under the [Modified MIT License](LICENSE).
---
## 7. Citation
```
@misc{kimiteam2025kimik2openagentic,
title={Kimi K2: Open Agentic Intelligence},
author={Kimi Team and Yifan Bai and Yiping Bao and Guanduo Chen and Jiahao Chen and Ningxin Chen and Ruijue Chen and Yanru Chen and Yuankun Chen and Yutian Chen and Zhuofu Chen and Jialei Cui and Hao Ding and Mengnan Dong and Angang Du and Chenzhuang Du and Dikang Du and Yulun Du and Yu Fan and Yichen Feng and Kelin Fu and Bofei Gao and Hongcheng Gao and Peizhong Gao and Tong Gao and Xinran Gu and Longyu Guan and Haiqing Guo and Jianhang Guo and Hao Hu and Xiaoru Hao and Tianhong He and Weiran He and Wenyang He and Chao Hong and Yangyang Hu and Zhenxing Hu and Weixiao Huang and Zhiqi Huang and Zihao Huang and Tao Jiang and Zhejun Jiang and Xinyi Jin and Yongsheng Kang and Guokun Lai and Cheng Li and Fang Li and Haoyang Li and Ming Li and Wentao Li and Yanhao Li and Yiwei Li and Zhaowei Li and Zheming Li and Hongzhan Lin and Xiaohan Lin and Zongyu Lin and Chengyin Liu and Chenyu Liu and Hongzhang Liu and Jingyuan Liu and Junqi Liu and Liang Liu and Shaowei Liu and T. Y. Liu and Tianwei Liu and Weizhou Liu and Yangyang Liu and Yibo Liu and Yiping Liu and Yue Liu and Zhengying Liu and Enzhe Lu and Lijun Lu and Shengling Ma and Xinyu Ma and Yingwei Ma and Shaoguang Mao and Jie Mei and Xin Men and Yibo Miao and Siyuan Pan and Yebo Peng and Ruoyu Qin and Bowen Qu and Zeyu Shang and Lidong Shi and Shengyuan Shi and Feifan Song and Jianlin Su and Zhengyuan Su and Xinjie Sun and Flood Sung and Heyi Tang and Jiawen Tao and Qifeng Teng and Chensi Wang and Dinglu Wang and Feng Wang and Haiming Wang and Jianzhou Wang and Jiaxing Wang and Jinhong Wang and Shengjie Wang and Shuyi Wang and Yao Wang and Yejie Wang and Yiqin Wang and Yuxin Wang and Yuzhi Wang and Zhaoji Wang and Zhengtao Wang and Zhexu Wang and Chu Wei and Qianqian Wei and Wenhao Wu and Xingzhe Wu and Yuxin Wu and Chenjun Xiao and Xiaotong Xie and Weimin Xiong and Boyu Xu and Jing Xu and Jinjing Xu and L. H. Xu and Lin Xu and Suting Xu and Weixin Xu and Xinran Xu and Yangchuan Xu and Ziyao Xu and Junjie Yan and Yuzi Yan and Xiaofei Yang and Ying Yang and Zhen Yang and Zhilin Yang and Zonghan Yang and Haotian Yao and Xingcheng Yao and Wenjie Ye and Zhuorui Ye and Bohong Yin and Longhui Yu and Enming Yuan and Hongbang Yuan and Mengjie Yuan and Haobing Zhan and Dehao Zhang and Hao Zhang and Wanlu Zhang and Xiaobin Zhang and Yangkun Zhang and Yizhi Zhang and Yongting Zhang and Yu Zhang and Yutao Zhang and Yutong Zhang and Zheng Zhang and Haotian Zhao and Yikai Zhao and Huabin Zheng and Shaojie Zheng and Jianren Zhou and Xinyu Zhou and Zaida Zhou and Zhen Zhu and Weiyu Zhuang and Xinxing Zu},
year={2025},
eprint={2507.20534},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2507.20534},
}
```
---
## 8. Contact Us
If you have any questions or concerns, please reach out to us at [support@moonshot.cn](mailto:support@moonshot.cn).
================================================
FILE: docs/deploy_guidance.md
================================================
# Kimi-K2 Deployment Guide
> [!Note]
> This guide only provides some examples of deployment commands for Kimi-K2, which may not be the optimal configuration. Since inference engines are still being updated frequently, please continue to follow the guidance from their homepage if you want to achieve better inference performance.
## vLLM Deployment
vLLM version v0.10.0rc1 or later is required.
The smallest deployment unit for Kimi-K2 FP8 weights with 128k seqlen on mainstream H200 or H20 platform is a cluster with 16 GPUs with either Tensor Parallel (TP) or "data parallel + expert parallel" (DP+EP).
Running parameters for this environment are provided below. You may scale up to more nodes and increase expert-parallelism to enlarge the inference batch size and overall throughput.
### Tensor Parallelism
When the parallelism degree ≤ 16, you can run inference with pure Tensor Parallelism. A sample launch command is:
``` bash
# start ray on node 0 and node 1
# node 0:
vllm serve $MODEL_PATH \
--port 8000 \
--served-model-name kimi-k2 \
--trust-remote-code \
--tensor-parallel-size 16 \
--enable-auto-tool-choice \
--tool-call-parser kimi_k2
```
**Key parameter notes:**
- `--tensor-parallel-size 16`: If using more than 16 GPUs, combine with pipeline-parallelism.
- `--enable-auto-tool-choice`: Required when enabling tool usage.
- `--tool-call-parser kimi_k2`: Required when enabling tool usage.
### Data Parallelism + Expert Parallelism
You can install libraries like DeepEP and DeepGEMM as needed. Then run the command (example on H200):
``` bash
# node 0
vllm serve $MODEL_PATH --port 8000 --served-model-name kimi-k2 --trust-remote-code --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-address $MASTER_IP --data-parallel-rpc-port $PORT --enable-expert-parallel --max-num-batched-tokens 8192 --max-num-seqs 256 --gpu-memory-utilization 0.85 --enable-auto-tool-choice --tool-call-parser kimi_k2
# node 1
vllm serve $MODEL_PATH --headless --data-parallel-start-rank 8 --port 8000 --served-model-name kimi-k2 --trust-remote-code --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-address $MASTER_IP --data-parallel-rpc-port $PORT --enable-expert-parallel --max-num-batched-tokens 8192 --max-num-seqs 256 --gpu-memory-utilization 0.85 --enable-auto-tool-choice --tool-call-parser kimi_k2
```
## SGLang Deployment
Similarly, we can use TP or DP+EP in SGLang for Deployment, here are the examples.
### Tensor Parallelism
Here is the simple example code to run TP16 with two nodes on H200:
``` bash
# Node 0
python -m sglang.launch_server --model-path $MODEL_PATH --tp 16 --dist-init-addr $MASTER_IP:50000 --nnodes 2 --node-rank 0 --trust-remote-code --tool-call-parser kimi_k2
# Node 1
python -m sglang.launch_server --model-path $MODEL_PATH --tp 16 --dist-init-addr $MASTER_IP:50000 --nnodes 2 --node-rank 1 --trust-remote-code --tool-call-parser kimi_k2
```
**Key parameter notes:**
- `--tool-call-parser kimi_k2`: Required when enabling tool usage.
### Data Parallelism + Expert Parallelism
Here is an example for large scale Prefill-Decode Disaggregation (4P12D H200) with DP+EP in SGLang:
``` bash
# for prefill node
MC_TE_METRIC=true SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL=10000000 SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=100000 SGLANG_DISAGGREGATION_WAITING_TIMEOUT=100000 PYTHONUNBUFFERED=1 \
python -m sglang.launch_server --model-path $MODEL_PATH \
--trust-remote-code --disaggregation-mode prefill --dist-init-addr $PREFILL_NODE0$:5757 --tp-size 32 --dp-size 32 --enable-dp-attention --host $LOCAL_IP --decode-log-interval 1 --disable-radix-cache --enable-deepep-moe --moe-dense-tp-size 1 --enable-dp-lm-head --disable-shared-experts-fusion --watchdog-timeout 1000000 --enable-two-batch-overlap --disaggregation-ib-device $IB_DEVICE --chunked-prefill-size 131072 --mem-fraction-static 0.85 --deepep-mode normal --ep-dispatch-algorithm dynamic --eplb-algorithm deepseek --max-running-requests 1024 --nnodes 4 --node-rank $RANK --tool-call-parser kimi_k2
# for decode node
SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=480 MC_TE_METRIC=true SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL=10000000 SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=100000 SGLANG_DISAGGREGATION_WAITING_TIMEOUT=100000 PYTHONUNBUFFERED=1 \
python -m sglang.launch_server --model-path $MODEL_PATH --trust-remote-code --disaggregation-mode decode --dist-init-addr $DECODE_NODE0:5757 --tp-size 96 --dp-size 96 --enable-dp-attention --host $LOCAL_IP --decode-log-interval 1 --context-length 2176 --disable-radix-cache --enable-deepep-moe --moe-dense-tp-size 1 --enable-dp-lm-head --disable-shared-experts-fusion --watchdog-timeout 1000000 --enable-two-batch-overlap --disaggregation-ib-device $IB_DEVICE --deepep-mode low_latency --mem-fraction-static 0.8 --cuda-graph-bs 480 --max-running-requests 46080 --ep-num-redundant-experts 96 --nnodes 12 --node-rank $RANK --tool-call-parser kimi_k2
# pdlb
PYTHONUNBUFFERED=1 python -m sglang.srt.disaggregation.launch_lb --prefill http://${PREFILL_NODE0}:30000 --decode http://${DECODE_NODE0}:30000
```
## KTransformers Deployment
Please copy all configuration files (i.e., everything except the .safetensors files) into the GGUF checkpoint folder at /path/to/K2. Then run:
``` bash
python ktransformers/server/main.py --model_path /path/to/K2 --gguf_path /path/to/K2 --cache_lens 30000
```
To enable AMX optimization, run:
``` bash
python ktransformers/server/main.py --model_path /path/to/K2 --gguf_path /path/to/K2 --cache_lens 30000 --optimize_config_path ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-fp8-linear-ggml-experts-serve-amx.yaml
```
## TensorRT-LLM Deployment
### Prerequisite
Please refer to [this guide](https://nvidia.github.io/TensorRT-LLM/installation/build-from-source-linux.html) to build TensorRT-LLM v1.0.0-rc2 from source and start a TRT-LLM docker container.
install blobfile by:
```bash
pip install blobfile
```
### Multi-node Serving
TensorRT-LLM supports multi-node inference. You can use mpirun to launch Kimi-K2 with multi-node jobs. We will use two nodes for this example.
#### mpirun
mpirun requires each node to have passwordless ssh access to the other node. We need to setup the environment inside the docker container. Run the container with host network and mount the current directory as well as model directory to the container.
```bash
# use host network
IMAGE=
NAME=test_2node_docker
# host1
docker run -it --name ${NAME}_host1 --ipc=host --gpus=all --network host --privileged --ulimit memlock=-1 --ulimit stack=67108864 -v ${PWD}:/workspace -v :/models/DeepSeek-V3 -w /workspace ${IMAGE}
# host2
docker run -it --name ${NAME}_host2 --ipc=host --gpus=all --network host --privileged --ulimit memlock=-1 --ulimit stack=67108864 -v ${PWD}:/workspace -v :/models/DeepSeek-V3 -w /workspace ${IMAGE}
```
Set up ssh inside the container
```bash
apt-get update && apt-get install -y openssh-server
# modify /etc/ssh/sshd_config
PermitRootLogin yes
PubkeyAuthentication yes
# modify /etc/ssh/sshd_config, change default port 22 to another unused port
port 2233
# modify /etc/ssh
```
Generate ssh key on host1 and copy to host2, vice versa.
```bash
# on host1
ssh-keygen -t ed25519 -f ~/.ssh/id_ed25519
ssh-copy-id -i ~/.ssh/id_ed25519.pub root@
# on host2
ssh-keygen -t ed25519 -f ~/.ssh/id_ed25519
ssh-copy-id -i ~/.ssh/id_ed25519.pub root@
# restart ssh service on host1 and host2
service ssh restart # or
/etc/init.d/ssh restart # or
systemctl restart ssh
```
Generate additional config for trtllm serve.
```bash
cat >/path/to/TensorRT-LLM/extra-llm-api-config.yml <:8,:8 \
-mca plm_rsh_args "-p 2233" \
--allow-run-as-root \
trtllm-llmapi-launch trtllm-serve serve \
--backend pytorch \
--tp_size 16 \
--ep_size 8 \
--kv_cache_free_gpu_memory_fraction 0.95 \
--trust_remote_code \
--max_batch_size 128 \
--max_num_tokens 4096 \
--extra_llm_api_options /path/to/TensorRT-LLM/extra-llm-api-config.yml \
--port 8000 \
```
## Others
Kimi-K2 reuses the `DeepSeekV3CausalLM` architecture and convert it's weight into proper shape to save redevelopment effort. To let inference engines distinguish it from DeepSeek-V3 and apply the best optimizations, we set `"model_type": "kimi_k2"` in `config.json`.
If you are using a framework that is not on the recommended list, you can still run the model by manually changing `model_type` to "deepseek_v3" in `config.json` as a temporary workaround. You may need to manually parse tool calls in case no tool call parser is available in your framework.
================================================
FILE: docs/tool_call_guidance.md
================================================
## Tool Calling
To enable the tool calling feature, you may need to set certain tool calling parser options when starting the service. See [deploy_guidance](./deploy_guidance.md) for details.
In Kimi-K2, a tool calling process includes:
- Passing function descriptions to Kimi-K2
- Kimi-K2 decides to make a function call and returns the necessary information for the function call to the user
- The user performs the function call, collects the call results, and passes the function call results to Kimi-K2
- Kimi-K2 continues to generate content based on the function call results until the model believes it has obtained sufficient information to respond to the user
### Preparing Tools
Suppose we have a function `get_weather` that can query the weather conditions in real-time.
This function accepts a city name as a parameter and returns the weather conditions. We need to prepare a structured description for it so that Kimi-K2 can understand its functionality.
```python
def get_weather(city):
return {"weather": "Sunny"}
# Collect the tool descriptions in tools
tools = [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Get weather information. Call this tool when the user needs to get weather information",
"parameters": {
"type": "object",
"required": ["city"],
"properties": {
"city": {
"type": "string",
"description": "City name",
}
}
}
}
}]
# Tool name->object mapping for easy calling later
tool_map = {
"get_weather": get_weather
}
```
### Chat with tools
We use `openai.OpenAI` to send messages to Kimi-K2 along with tool descriptions. Kimi-K2 will autonomously decide whether to use and how to use the provided tools.
If Kimi-K2 believes a tool call is needed, it will return a result with `finish_reason='tool_calls'`. At this point, the returned result includes the tool call information.
After calling tools with the provided information, we then need to append the tool call results to the chat history and continue calling Kimi-K2.
Kimi-K2 may need to call tools multiple times until the model believes the current results can answer the user's question. We should check `finish_reason` until it is not `tool_calls`.
The results obtained by the user after calling the tools should be added to `messages` with `role='tool'`.
```python
import json
from openai import OpenAI
model_name='moonshotai/Kimi-K2-Instruct'
client = OpenAI(base_url=endpoint,
api_key='xxx')
messages = [
{"role": "user", "content": "What's the weather like in Beijing today? Let's check using the tool."}
]
finish_reason = None
while finish_reason is None or finish_reason == "tool_calls":
completion = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.3,
tools=tools,
tool_choice="auto",
)
choice = completion.choices[0]
finish_reason = choice.finish_reason
# Note: The finish_reason when tool calls end may vary across different engines, so this condition check needs to be adjusted accordingly
if finish_reason == "tool_calls":
messages.append(choice.message)
for tool_call in choice.message.tool_calls:
tool_call_name = tool_call.function.name
tool_call_arguments = json.loads(tool_call.function.arguments)
tool_function = tool_map[tool_call_name]
tool_result = tool_function(tool_call_arguments)
print("tool_result", tool_result)
messages.append({
"role": "tool",
"tool_call_id": tool_call.id,
"name": tool_call_name,
"content": json.dumps(tool_result),
})
print('-' * 100)
print(choice.message.content)
```
### Tool Calling in Streaming Mode
Tool calling can also be used in streaming mode. In this case, we need to collect the tool call information returned in the stream until we have a complete tool call. Please refer to the code below:
```python
messages = [
{"role": "user", "content": "What's the weather like in Beijing today? Let's check using the tool."}
]
finish_reason = None
msg = ''
while finish_reason is None or finish_reason == "tool_calls":
completion = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.3,
tools=tools,
tool_choice="auto",
stream=True
)
tool_calls = []
for chunk in completion:
delta = chunk.choices[0].delta
if delta.content:
msg += delta.content
if delta.tool_calls:
for tool_call_chunk in delta.tool_calls:
if tool_call_chunk.index is not None:
# Extend the tool_calls list
while len(tool_calls) <= tool_call_chunk.index:
tool_calls.append({
"id": "",
"type": "function",
"function": {
"name": "",
"arguments": ""
}
})
tc = tool_calls[tool_call_chunk.index]
if tool_call_chunk.id:
tc["id"] += tool_call_chunk.id
if tool_call_chunk.function.name:
tc["function"]["name"] += tool_call_chunk.function.name
if tool_call_chunk.function.arguments:
tc["function"]["arguments"] += tool_call_chunk.function.arguments
finish_reason = chunk.choices[0].finish_reason
# Note: The finish_reason when tool calls end may vary across different engines, so this condition check needs to be adjusted accordingly
if finish_reason == "tool_calls":
for tool_call in tool_calls:
tool_call_name = tool_call['function']['name']
tool_call_arguments = json.loads(tool_call['function']['arguments'])
tool_function = tool_map[tool_call_name]
tool_result = tool_function(tool_call_arguments)
messages.append({
"role": "tool",
"tool_call_id": tool_call['id'],
"name": tool_call_name,
"content": json.dumps(tool_result),
})
# The text generated by the tool call is not the final version, reset msg
msg = ''
print(msg)
```
### Manually Parsing Tool Calls
The tool call requests generated by Kimi-K2 can also be parsed manually, which is especially useful when the service you are using does not provide a tool-call parser.
The tool call requests generated by Kimi-K2 are wrapped by `<|tool_calls_section_begin|>` and `<|tool_calls_section_end|>`,
with each tool call wrapped by `<|tool_call_begin|>` and `<|tool_call_end|>`. The tool ID and arguments are separated by `<|tool_call_argument_begin|>`.
The format of the tool ID is `functions.{func_name}:{idx}`, from which we can parse the function name.
Based on the above rules, we can directly post a request to the completions interface and manually parse tool calls.
```python
import requests
from transformers import AutoTokenizer
messages = [
{"role": "user", "content": "What's the weather like in Beijing today? Let's check using the tool."}
]
msg = ''
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
while True:
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
tools=tools,
add_generation_prompt=True,
)
payload = {
"model": model_name,
"prompt": text,
"max_tokens": 512
}
response = requests.post(
f"{endpoint}/completions",
headers={"Content-Type": "application/json"},
json=payload,
stream=False,
)
raw_out = response.json()
raw_output = raw_out["choices"][0]["text"]
tool_calls = extract_tool_call_info(raw_output)
if len(tool_calls) == 0:
# No tool calls
msg = raw_output
break
else:
for tool_call in tool_calls:
tool_call_name = tool_call['function']['name']
tool_call_arguments = json.loads(tool_call['function']['arguments'])
tool_function = tool_map[tool_call_name]
tool_result = tool_function(tool_call_arguments)
messages.append({
"role": "tool",
"tool_call_id": tool_call['id'],
"name": tool_call_name,
"content": json.dumps(tool_result),
})
print('-' * 100)
print(msg)
```
Here, `extract_tool_call_info` parses the model output and returns the model call information. A simple implementation would be:
```python
def extract_tool_call_info(tool_call_rsp: str):
if '<|tool_calls_section_begin|>' not in tool_call_rsp:
# No tool calls
return []
import re
pattern = r"<\|tool_calls_section_begin\|>(.*?)<\|tool_calls_section_end\|>"
tool_calls_sections = re.findall(pattern, tool_call_rsp, re.DOTALL)
# Extract multiple tool calls
func_call_pattern = r"<\|tool_call_begin\|>\s*(?P[\w\.]+:\d+)\s*<\|tool_call_argument_begin\|>\s*(?P.*?)\s*<\|tool_call_end\|>"
tool_calls = []
for match in re.findall(func_call_pattern, tool_calls_sections[0], re.DOTALL):
function_id, function_args = match
# function_id: functions.get_weather:0
function_name = function_id.split('.')[1].split(':')[0]
tool_calls.append(
{
"id": function_id,
"type": "function",
"function": {
"name": function_name,
"arguments": function_args
}
}
)
return tool_calls
```
