Qwen2.5-1.5B-Instruct-GPTQ-Int4

このバージョンの Qwen2.5-1.5B-Instruct-GPTQ-Int4 は、w4a16 量子化を用いて Axera NPU 上で動作するように変換されています。

このモデルは以下の LoRA で最適化されています:

対応する Pulsar2 バージョン: 3.4(未リリース)

変換ツールのリンク

モデル変換に興味のある方は、元のリポジトリから axmodel をエクスポートしてみてください: https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GPTQ-Int4

Pulsar2 リンク: Huggingface の LLM を axmodel に変換する方法

AXera NPU LLM Runtime

対応プラットフォーム

• AX650
  • AX650N DEMO Board
  • M4N-Dock(爱芯派Pro)
  • M.2 Accelerator card
• AX630C
  • 開発中

使い方

このリポジトリのすべてのファイルをデバイスにダウンロードします

root@ax650:/mnt/qtang/llm-test/qwen2.5-1.5b# tree -L 1
.
├── qwen2.5-1.5b-gptq-int4-ax650
├── qwen2.5_tokenizer
├── qwen2.5_tokenizer.py
├── main_axcl_aarch64
├── main_axcl_x86
├── main_prefill
├── post_config.json
├── run_qwen2.5_1.5b_gptq_int4_ax650.sh
├── run_qwen2.5_1.5b_gptq_int4_axcl_aarch64.sh
└── run_qwen2.5_1.5b_gptq_int4_axcl_x86.sh

Tokenizer サービスの起動

root@ax650:/mnt/qtang/llm-test/qwen2.5-1.5b# python3 qwen2.5_tokenizer.py --port 12345
None None 151645 <|im_end|>
<|im_start|>system
You are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>
<|im_start|>user
hello world<|im_end|>
<|im_start|>assistant

[151644, 8948, 198, 2610, 525, 1207, 16948, 11, 3465, 553, 54364, 14817, 13, 1446, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 14990, 1879, 151645, 198, 151644, 77091, 198]
http://localhost:12345

AX650 ホスト(M4N-Dock(爱芯派Pro)や AX650N DEMO Board など)での推論

別のターミナルを開き、run_qwen2.5_1.5b_gptq_int4_ax650.sh を実行します

root@ax650:/mnt/qtang/llm-test/qwen2.5-1.5b# ./run_qwen2.5_1.5b_gptq_int4_ax650.sh
[I][                            Init][ 125]: LLM init start
bos_id: -1, eos_id: 151645
100% | ████████████████████████████████ |  31 /  31 [1.62s<1.62s, 19.14 count/s] init post axmodel ok,remain_cmm(2731 MB)
[I][                            Init][ 241]: max_token_len : 1023
[I][                            Init][ 246]: kv_cache_size : 256, kv_cache_num: 1023
[I][                            Init][ 254]: prefill_token_num : 128
[I][                     load_config][ 281]: load config:
{
    "enable_repetition_penalty": false,
    "enable_temperature": true,
    "enable_top_k_sampling": true,
    "enable_top_p_sampling": false,
    "penalty_window": 20,
    "repetition_penalty": 1.2,
    "temperature": 0.9,
    "top_k": 10,
    "top_p": 0.8
}

[I][                            Init][ 268]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running
>> who are you
I am Qwen, a large language model created by Alibaba Cloud. I assist people with language translation and o                                                                       ther tasks by understanding human language and generating more appropriate words, phrases, or sentences.

[N][                             Run][ 610]: hit eos,avg 18.68 token/s

>> 1+1=?
1+1 equals 2.

[N][                             Run][ 610]: hit eos,avg 19.43 token/s

M.2 アクセラレータカードでの推論

M.2 アクセラレータカードとは? Raspberry Pi 5 を使ったデモを示します。

(base) axera@raspberrypi:~/samples/qwen2.5-1.5b $ ./run_qwen2.5_1.5b_gptq_int4_axcl_aarch64.sh
build time: Feb 13 2025 15:44:57
[I][                            Init][ 111]: LLM init start
bos_id: -1, eos_id: 151645
100% | ████████████████████████████████ |  31 /  31 [22.80s<22.80s, 1.36 count/s] init post axmodel okremain_cmm(6219 MB)
[I][                            Init][ 226]: max_token_len : 1023
[I][                            Init][ 231]: kv_cache_size : 256, kv_cache_num: 1023
[I][                     load_config][ 282]: load config:
{
    "enable_repetition_penalty": false,
    "enable_temperature": true,
    "enable_top_k_sampling": true,
    "enable_top_p_sampling": false,
    "penalty_window": 20,
    "repetition_penalty": 1.2,
    "temperature": 0.9,
    "top_k": 10,
    "top_p": 0.8
}

[I][                            Init][ 288]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running

>> who are you
I am Qwen, an AI developed by Alibaba Cloud. I can help you find resources, answer questions, express myself, and interact with you in multiple ways such as text, speech, and image. Please feel free to let me know how I can assist you.

[N][                             Run][ 610]: hit eos,avg 15.36 token/s

>> q

(base) axera@raspberrypi:~/samples/qwen2.5-1.5b $ axcl-smi
+------------------------------------------------------------------------------------------------+
| AXCL-SMI  V2.26.0_20250205130139                                Driver  V2.26.0_20250205130139 |
+-----------------------------------------+--------------+---------------------------------------+
| Card  Name                     Firmware | Bus-Id       |                          Memory-Usage |
| Fan   Temp                Pwr:Usage/Cap | CPU      NPU |                             CMM-Usage |
|=========================================+==============+=======================================|
|    0  AX650N                    V2.26.0 | 0000:01:00.0 |                170 MiB /      945 MiB |
|   --   39C                      -- / -- | 0%        0% |               1053 MiB /     7040 MiB |
+-----------------------------------------+--------------+---------------------------------------+

+------------------------------------------------------------------------------------------------+
| Processes:                                                                                     |
| Card      PID  Process Name                                                   NPU Memory Usage |
|================================================================================================|
|    0    17325  /home/axera/samples/qwen2.5-1.5b/main_axcl_aarch64               1037736 KiB |
+------------------------------------------------------------------------------------------------+