pulsar2 llm_build --input_path Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4  \
                  --output_path Qwen/Qwen2.5-7B-Instruct-GPTQ-Int4-ctx-ax650 \
                  --hidden_state_type bf16 --kv_cache_len 2047 --prefill_len 128 \
                  --last_kv_cache_len 128 \
                  --last_kv_cache_len 256 \
                  --last_kv_cache_len 384 \
                  --last_kv_cache_len 512 \
                  --last_kv_cache_len 640 \
                  --last_kv_cache_len 768 \
                  --last_kv_cache_len 896 \
                  --last_kv_cache_len 1024 \
                  --chip AX650 -c 1 --parallel 8

Support Platform

AX650
- AX650N DEMO Board
- M4N-Dock(爱芯派Pro)
- M4N-HAT
- M.2 Accelerator card

Chips	w8a16	w4a16	DDR(w8)	Flash(w8)	DDR(w4)	Flash(w4)
AX650	2.8 tokens/sec	5.0 tokens/sec			5.2GB	5.7GB

How to use

Download all files from this repository to the device

(base) axera@raspberrypi:~/samples/AXERA-TECH/Qwen2.5-7B-Instruct $ tree -L 1
.
├── config.json
├── main_api
├── main_api_ax650
├── main_api_axcl_aarch64
├── main_api_axcl_x86
├── main_ax650
├── main_axcl_aarch64
├── main_axcl_x86
├── post_config.json
├── qwen2.5-7b-ctx-int4-ax650
├── qwen2.5_tokenizer
├── qwen2.5_tokenizer_uid.py
├── README.md
├── run_qwen2.5_7b_ctx_ax650.sh
├── run_qwen2.5_7b_ctx_int4_ax650.sh
├── run_qwen2.5_7b_ctx_int4_axcl_aarch64.sh
└── run_qwen2.5_7b_ctx_int4_axcl_x86.sh

3 directories, 15 files

Start the Tokenizer service

(axcl) axera@raspberrypi:~/samples/AXERA-TECH/Qwen2.5-7B-Instruct $ python qwen2.5_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

System prompt cache

The System prompt can be preset through the configuration file from --system_prompt
The System prompt can be cached in the form of kv cache to a specified folder for quick loading at the next run time from --kvcache_path
This folder needs to be created manually before running, for example mkdir kvcache

(base) axera@raspberrypi:~/samples/AXERA-TECH/Qwen2.5-7B-Instruct $ cat run_qwen2.5_7b_ctx_int4_axcl_aarch64.sh
./main_axcl_aarch64 \
--template_filename_axmodel "qwen2.5-7b-ctx-int4-ax650/qwen2_p128_l%d_together.axmodel" \
--axmodel_num 28 \
--url_tokenizer_model "http://0.0.0.0:12345" \
--filename_post_axmodel "qwen2.5-7b-ctx-int4-ax650/qwen2_post.axmodel" \
--filename_tokens_embed "qwen2.5-7b-ctx-int4-ax650/model.embed_tokens.weight.bfloat16.bin" \
--tokens_embed_num 152064 \
--tokens_embed_size 3584 \
--use_mmap_load_embed 1 \
--live_print 1 \
--devices 0


#--system_prompt "你的名字叫小智（allen）,你是一个人畜无害的AI助手。深圳市今天（4月1日）阴天，愚人节，气温在14°C至19°C之间，微风。" \
#--kvcache_path "./kvcache" \

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

TODO

Inference with M.2 Accelerator card

What is M.2 Accelerator card?, Show this DEMO based on Raspberry PI 5.

(base) axera@raspberrypi:~/samples/AXERA-TECH/Qwen2.5-7B-Instruct $ ./run_qwen2.5_7b_ctx_int4_axcl_aarch64.sh
[I][                            Init][ 130]: LLM init start
[I][                            Init][  34]: connect http://0.0.0.0:12345 ok
[I][                            Init][  57]: uid: ae9adea5-c64e-47df-92ca-29cbcc5a865f
bos_id: -1, eos_id: 151645
  3% | ██                                |   1 /  31 [0.49s<15.16s, 2.04 count/s] tokenizer init ok[I][                            Init][  45]: LLaMaEmbedSelector use mmap
  6% | ███                               |   2 /  31 [0.49s<7.59s, 4.08 count/s] embed_selector init ok
[I][                             run][  30]: AXCLWorker start with devid 0
 54% | ████████████████████████████ █ █  █ ██ ██               |  17 /  31 [39.92s<77.35s, 0.40 count/s] init 24 axmodel ok,devid(0) remain_cmm(-1 MB)            |  16 /  31 [39.92s<77.35s,100% | ████████████████████████████████ |  31 /  31 [80.60s<83.29s, 0.37 count/s] init post axmodel ok,remain_cmm(1324 MB)1891 MB)
[I][                            Init][ 221]: max_token_len : 2047
[I][                            Init][ 224]: kv_cache_size : 512, kv_cache_num: 2047
[I][                            Init][ 232]: prefill_token_num : 128
[I][                            Init][ 236]: grp: 1, prefill_max_token_num : 1
[I][                            Init][ 236]: grp: 2, prefill_max_token_num : 128
[I][                            Init][ 236]: grp: 3, prefill_max_token_num : 256
[I][                            Init][ 236]: grp: 4, prefill_max_token_num : 384
[I][                            Init][ 236]: grp: 5, prefill_max_token_num : 512
[I][                            Init][ 236]: grp: 6, prefill_max_token_num : 640
[I][                            Init][ 236]: grp: 7, prefill_max_token_num : 768
[I][                            Init][ 236]: grp: 8, prefill_max_token_num : 896
[I][                            Init][ 236]: grp: 9, prefill_max_token_num : 1024
[I][                            Init][ 240]: prefill_max_token_num : 1024
________________________
|    ID| remain cmm(MB)|
========================
|     0|           1324|
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
[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][ 263]: LLM init ok
Type "q" to exit, Ctrl+c to stop current running
[I][          GenerateKVCachePrefill][ 324]: input token num : 21, prefill_split_num : 1 prefill_grpid : 2
[I][          GenerateKVCachePrefill][ 367]: input_num_token:21
[I][                            main][ 234]: precompute_len: 21
[I][                            main][ 235]: system_prompt: You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
prompt >> nice
[I][                      SetKVCache][ 614]: prefill_grpid:2 kv_cache_num:128 precompute_len:21 input_num_token:9
[I][                      SetKVCache][ 617]: current prefill_max_token_num:896
[I][                             Run][ 855]: input token num : 9, prefill_split_num : 1
[I][                             Run][ 887]: input_num_token:9
[I][                             Run][1016]: ttft: 928.08 ms
Nice to meet you! If you have any questions or need some help, feel free to ask.

[N][                             Run][1168]: hit eos,avg 4.36 token/s

[I][                      GetKVCache][ 583]: precompute_len:50, remaining:974
prompt >> q
[I][                             run][  80]: AXCLWorker exit with devid 0
(base) axera@raspberrypi:~/samples/AXERA-TECH/Qwen2.5-7B-Instruct $