File size: 11,537 Bytes
98c3a60
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3-1.7B
pipeline_tag: text-generation
tags:
- Qwen
- Qwen3
- Int8
---


# Qwen3-1.7B-Int8

This version of Qwen3-1.7B-Int8 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: 4.0-temp(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/Qwen/Qwen3-1.7B

[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)
 
|Chips|w8a16|w4a16|
|--|--|--|
|AX650| 9.5 tokens/sec|TBD|

## How to use

Download all files from this repository to the device

```
root@ax650:/mnt/qtang/llm-test/qwen3-1.7b# tree -L 1
.
|-- config.json
|-- main_ax650
|-- main_axcl_aarch64
|-- main_axcl_x86
|-- post_config.json
|-- qwen2.5_tokenizer
|-- qwen3-1.7b-ax650
|-- qwen3_tokenizer
|-- qwen3_tokenizer_uid.py
|-- run_qwen3_1.7b_int8_ctx_ax650.sh
|-- run_qwen3_1.7b_int8_ctx_axcl_aarch64.sh
`-- run_qwen3_1.7b_int8_ctx_axcl_x86.sh

3 directories, 9 files
root@ax650:/mnt/qtang/llm-test/qwen3-1.7b#
```

#### Start the Tokenizer service

Install requirement

```
pip install transformers jinja2
```

```
root@ax650:/mnt/qtang/llm-test/qwen3-1.7b# python3 qwen3_tokenizer_uid.py
None of PyTorch, TensorFlow >= 2.0, or Flax have been found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
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_qwen3_1.7b_int8_ctx_ax650.sh`

```
root@ax650:/mnt/qtang/llm-test/qwen3-1.7b# ./run_qwen3_1.7b_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: 7a057c11-c513-485f-84a1-1d28dcbeb89d
bos_id: -1, eos_id: 151645
  3% | β–ˆβ–ˆ                                |   1 /  31 [3.97s<123.16s, 0.25 count/s] tokenizer init ok
[I][                            Init][  26]: LLaMaEmbedSelector use mmap
100% | β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ |  31 /  31 [23.76s<23.76s, 1.30 count/s] init post axmodel ok,remain_cmm(8740 MB)
[I][                            Init][ 188]: max_token_len : 2559
[I][                            Init][ 193]: kv_cache_size : 1024, kv_cache_num: 2559
[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 : 512
[I][                            Init][ 205]: grp: 3, prefill_max_token_num : 1024
[I][                            Init][ 205]: grp: 4, prefill_max_token_num : 1536
[I][                            Init][ 205]: grp: 5, prefill_max_token_num : 2048
[I][                            Init][ 209]: prefill_max_token_num : 2048
[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][ 270]: input token num : 21, prefill_split_num : 1 prefill_grpid : 2
[I][          GenerateKVCachePrefill][ 307]: input_num_token:21
[I][                            main][ 230]: precompute_len: 21
[I][                            main][ 231]: system_prompt: You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
prompt >> 1+1=?
[I][                      SetKVCache][ 530]: prefill_grpid:2 kv_cache_num:512 precompute_len:21 input_num_token:16
[I][                      SetKVCache][ 533]: current prefill_max_token_num:1920
[I][                             Run][ 659]: input token num : 16, prefill_split_num : 1
[I][                             Run][ 685]: input_num_token:16
[I][                             Run][ 808]: ttft: 678.72 ms
<think>

</think>

1 + 1 = 2.

[N][                             Run][ 922]: hit eos,avg 9.16 token/s

[I][                      GetKVCache][ 499]: precompute_len:49, remaining:1999
prompt >> who are you?
[I][                      SetKVCache][ 530]: prefill_grpid:2 kv_cache_num:512 precompute_len:49 input_num_token:16
[I][                      SetKVCache][ 533]: current prefill_max_token_num:1920
[I][                             Run][ 659]: input token num : 16, prefill_split_num : 1
[I][                             Run][ 685]: input_num_token:16
[I][                             Run][ 808]: ttft: 677.87 ms
<think>

</think>

I am Qwen, a large language model developed by Alibaba Cloud. I can answer questions,
help with tasks, and provide information on various topics. I am designed to be helpful and useful to users.

[N][                             Run][ 922]: hit eos,avg 9.13 token/s

[I][                      GetKVCache][ 499]: precompute_len:110, remaining:1938
prompt >> q
```

#### Inference with M.2 Accelerator card

[What is M.2 Accelerator card?](https://axcl-docs.readthedocs.io/zh-cn/latest/doc_guide_hardware.html), Show this DEMO based on Raspberry PI 5.

```
(base) axera@raspberrypi:~/samples/qwen3-1.7b $ ./run_qwen3_1.7b_int8_ctx_axcl_aarch64.sh
[I][                            Init][ 136]: LLM init start
[I][                            Init][  34]: connect http://127.0.0.1:12345 ok
[I][                            Init][  57]: uid: ea509ef6-ab6c-49b0-9dcf-931db2ce1bf7
bos_id: -1, eos_id: 151645
  3% | β–ˆβ–ˆ                                |   1 /  31 [0.98s<30.47s, 1.02 count/s] tokenizer init ok
[I][                            Init][  45]: LLaMaEmbedSelector use mmap
  6% | β–ˆβ–ˆβ–ˆ                               |   2 /  31 [0.98s<15.24s, 2.03 count/s] embed_selector init ok
[I][                             run][  30]: AXCLWorker start with devid 0
100% | β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ |  31 /  31 [49.40s<49.40s, 0.63 count/s] init post axmodel ok,remain_cmm(3788 MB)
[I][                            Init][ 237]: max_token_len : 2559
[I][                            Init][ 240]: kv_cache_size : 1024, kv_cache_num: 2559
[I][                            Init][ 248]: prefill_token_num : 128
[I][                            Init][ 252]: grp: 1, prefill_max_token_num : 1
[I][                            Init][ 252]: grp: 2, prefill_max_token_num : 512
[I][                            Init][ 252]: grp: 3, prefill_max_token_num : 1024
[I][                            Init][ 252]: grp: 4, prefill_max_token_num : 1536
[I][                            Init][ 252]: grp: 5, prefill_max_token_num : 2048
[I][                            Init][ 256]: prefill_max_token_num : 2048
________________________
|    ID| remain cmm(MB)|
========================
|     0|           3788|
Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―Β―
[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][ 279]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running
[I][          GenerateKVCachePrefill][ 335]: input token num : 21, prefill_split_num : 1 prefill_grpid : 2
[I][          GenerateKVCachePrefill][ 372]: input_num_token:21
[I][                            main][ 236]: precompute_len: 21
[I][                            main][ 237]: system_prompt: You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
prompt >> 1+2=?
[I][                      SetKVCache][ 628]: prefill_grpid:2 kv_cache_num:512 precompute_len:21 input_num_token:16
[I][                      SetKVCache][ 631]: current prefill_max_token_num:1920
[I][                             Run][ 869]: input token num : 16, prefill_split_num : 1
[I][                             Run][ 901]: input_num_token:16
[I][                             Run][1030]: ttft: 796.97 ms
<think>

</think>

1 + 2 = 3.

[N][                             Run][1182]: hit eos,avg 7.43 token/s

[I][                      GetKVCache][ 597]: precompute_len:49, remaining:1999
prompt >> who are you?
[I][                      SetKVCache][ 628]: prefill_grpid:2 kv_cache_num:512 precompute_len:49 input_num_token:16
[I][                      SetKVCache][ 631]: current prefill_max_token_num:1920
[I][                             Run][ 869]: input token num : 16, prefill_split_num : 1
[I][                             Run][ 901]: input_num_token:16
[I][                             Run][1030]: ttft: 800.01 ms
<think>

</think>

I am Qwen, a large language model developed by Alibaba Cloud. I can help with various tasks,
such as answering questions, writing text, providing explanations, and more. If you have any questions or need assistance, feel free to ask!

[N][                             Run][1182]: hit eos,avg 7.42 token/s

[I][                      GetKVCache][ 597]: precompute_len:118, remaining:1930
prompt >> q
[I][                             run][  80]: AXCLWorker exit with devid 0
(base) axera@raspberrypi:~/samples/qwen3-1.7b $
(base) axera@raspberrypi:~ $ axcl-smi
+------------------------------------------------------------------------------------------------+
| AXCL-SMI  V3.4.0_20250423020139                                  Driver  V3.4.0_20250423020139 |
+-----------------------------------------+--------------+---------------------------------------+
| Card  Name                     Firmware | Bus-Id       |                          Memory-Usage |
| Fan   Temp                Pwr:Usage/Cap | CPU      NPU |                             CMM-Usage |
|=========================================+==============+=======================================|
|    0  AX650N                     V3.4.0 | 0000:01:00.0 |                183 MiB /      945 MiB |
|   --   38C                      -- / -- | 0%        0% |               3251 MiB /     7040 MiB |
+-----------------------------------------+--------------+---------------------------------------+

+------------------------------------------------------------------------------------------------+
| Processes:                                                                                     |
| Card      PID  Process Name                                                   NPU Memory Usage |
|================================================================================================|
|    0    71266  /home/axera/samples/qwen3-1.7b/main_axcl_aarch64                    2193524 KiB |
+------------------------------------------------------------------------------------------------+
(base) axera@raspberrypi:~ $ 
```