---
license: mit
language:
- en
base_model:
- openbmb/MiniCPM4-0.5B
pipeline_tag: text-generation
tags:
- minicpm4
- int8
---


# MiniCPM4-0.5B-Int8

This version of MiniCPM4-0.5B has been converted to run on the Axera NPU using **w8a16** quantization.

This model has been optimized with the following LoRA: 

Compatible with Pulsar2 version: 4.2(Not released yet)

## Convert tools links:

For those who are interested in model conversion, you can try to export axmodel through the original repo :
https://huggingface.co/openbmb/MiniCPM4-0.5B

[Pulsar2 Link, How to Convert LLM from Huggingface to axmodel](https://pulsar2-docs.readthedocs.io/en/latest/appendix/build_llm.html) 

[AXera NPU LLM Runtime](https://github.com/AXERA-TECH/ax-llm) 


## Support Platform

- AX650
  - [M4N-Dock(爱芯派Pro)](https://wiki.sipeed.com/hardware/zh/maixIV/m4ndock/m4ndock.html)
  - [M.2 Accelerator card](https://axcl-docs.readthedocs.io/zh-cn/latest/doc_guide_hardware.html)
- AX630C
  - [爱芯派2](https://axera-pi-2-docs-cn.readthedocs.io/zh-cn/latest/index.html)
  - [Module-LLM](https://docs.m5stack.com/zh_CN/module/Module-LLM)
  - [LLM630 Compute Kit](https://docs.m5stack.com/zh_CN/core/LLM630%20Compute%20Kit)
 
|Chips|w8a16|w4a16|
|--|--|--|
|AX650| 36 tokens/sec|TBD|
|AX630C| 12 tokens/sec|TBD|

## How to use

Download all files from this repository to the device

```
root@ax650:/mnt/qtang/llm-test/minicpm4-0.5b-ctx# tree -L 1
.
|-- main_ax650
|-- main_axcl_aarch64
|-- main_axcl_x86
|-- minicpm4-0.5b-int8-ctx-ax650
|-- minicpm4_tokenizer
|-- minicpm4_tokenizer_uid.py
|-- post_config.json
|-- run_minicpm4_0.5b_int8_ctx_ax650.sh
`-- run_minicpm4_0.5b_int8_ctx_axcl_x86.sh
2 directories, 7 files
```

#### Start the Tokenizer service

Install requirement

```
pip install transformers jinja2
```

```
root@ax650:/mnt/qtang/llm-test/minicpm4-0.5b-ctx# python3 minicpm4_tokenizer_uid.py
Server running at http://0.0.0.0:12345
```

#### Inference with AX650 Host, such as M4N-Dock(爱芯派Pro) or AX650N DEMO Board

Open another terminal and run `run_minicpm4_0.5b_int8_ctx_ax650.sh`

```
root@ax650:/mnt/qtang/llm-test/minicpm4-0.5b-ctx# ./run_minicpm4_0.5b_int8_ctx_ax650.sh
[I][                            Init][ 110]: LLM init start
[I][                            Init][  34]: connect http://127.0.0.1:12345 ok
[I][                            Init][  57]: uid: c779ded0-ff14-4877-869b-1aacc948f2d8
bos_id: 1, eos_id: 73440
100% | ████████████████████████████████ |  27 /  27 [2.53s<2.53s, 10.67 count/s] init post axmodel ok,remain_cmm(4244 MB)
[I][                            Init][ 188]: max_token_len : 1023
[I][                            Init][ 193]: kv_cache_size : 128, kv_cache_num: 1023
[I][                            Init][ 201]: prefill_token_num : 128
[I][                            Init][ 205]: grp: 1, prefill_max_token_num : 1
[I][                            Init][ 205]: grp: 2, prefill_max_token_num : 128
[I][                            Init][ 205]: grp: 3, prefill_max_token_num : 512
[I][                            Init][ 209]: prefill_max_token_num : 512
[I][                     load_config][ 282]: load config:
{
    "enable_repetition_penalty": false,
    "enable_temperature": false,
    "enable_top_k_sampling": true,
    "enable_top_p_sampling": false,
    "penalty_window": 20,
    "repetition_penalty": 1.2,
    "temperature": 0.9,
    "top_k": 1,
    "top_p": 0.8
}

[I][                            Init][ 218]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running
[I][          GenerateKVCachePrefill][ 271]: input token num : 25, prefill_split_num : 1 prefill_grpid : 2
[I][          GenerateKVCachePrefill][ 308]: input_num_token:25
[I][                            main][ 230]: precompute_len: 25
[I][                            main][ 231]: system_prompt: You are MiniCPM4, created by ModelBest. You are a helpful assistant.
prompt >> 你是谁?
[I][                      SetKVCache][ 531]: prefill_grpid:2 kv_cache_num:128 precompute_len:25 input_num_token:12
[I][                      SetKVCache][ 534]: current prefill_max_token_num:384
[I][                             Run][ 660]: input token num : 12, prefill_split_num : 1
[I][                             Run][ 686]: input_num_token:12
[I][                             Run][ 829]: ttft: 147.65 ms
你好，我是MiniCPM系列模型，由面壁智能和OpenBMB开源社区开发。详细信息请访问https://github.com/OpenBMB/

[N][                             Run][ 943]: hit eos,avg 35.75 token/s

[I][                      GetKVCache][ 500]: precompute_len:162, remaining:350
prompt >> 9.9与9.11
[I][                      SetKVCache][ 531]: prefill_grpid:3 kv_cache_num:512 precompute_len:162 input_num_token:17
[I][                      SetKVCache][ 534]: current prefill_max_token_num:256
[I][                             Run][ 660]: input token num : 17, prefill_split_num : 1
[I][                             Run][ 686]: input_num_token:17
[I][                             Run][ 829]: ttft: 274.38 ms
9.9比9.11大。

[N][                             Run][ 943]: hit eos,avg 35.44 token/s

[I][                      GetKVCache][ 500]: precompute_len:189, remaining:323
prompt >> q
root@ax650:/mnt/qtang/llm-test/minicpm4-0.5b-ctx#
```