FastVLM-1.5B-GPTQ-Int4

FastVLM-1.5B-GPTQ-Int4

This version of FastVLM-1.5B-GPTQ-Int4 has been converted to run on the Axera NPU using w4a16 quantization.

This model has been optimized with the following LoRA:

Compatible with Pulsar2 version: 5.1-patch1.

Please note that the context of the model is 1k and the maximum prefill length is 640 tokens.

Convert tools links:

For those who are interested in model conversion, you can try to quant and export axmodel through the original repo:

https://huggingface.co/apple/FastVLM-1.5B

How to Convert LLM from Huggingface to axmodel[TODO]

Support Platform

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

How to use

安装 axllm

方式一：克隆仓库后执行安装脚本：

git clone -b axllm https://github.com/AXERA-TECH/ax-llm.git
cd ax-llm
./install.sh

方式二：一行命令安装（默认分支 axllm）：

curl -fsSL https://raw.githubusercontent.com/AXERA-TECH/ax-llm/axllm/install.sh | bash

方式三：下载Github Actions CI 导出的可执行程序（适合没有编译环境的用户）：

如果没有编译环境，请到： https://github.com/AXERA-TECH/ax-llm/actions?query=branch%3Aaxllm 下载 最新 CI 导出的可执行程序（axllm），然后：

chmod +x axllm
sudo mv axllm /usr/bin/axllm

模型下载（Hugging Face）

先创建模型目录并进入，然后下载到该目录：

mkdir -p AXERA-TECH/FastVLM-1.5B-GPTQ-Int4
cd AXERA-TECH/FastVLM-1.5B-GPTQ-Int4
hf download AXERA-TECH/FastVLM-1.5B-GPTQ-Int4 --local-dir .

# structure of the downloaded files
tree -L 3
.
`-- AXERA-TECH
    `-- FastVLM-1.5B-GPTQ-Int4
        |-- FastVLM_tokenizer.txt
        |-- README.md
        |-- config.json
        |-- image.png
        |-- image_encoder_1024x1024.axmodel
        |-- image_encoder_512x512.axmodel
        |-- llava_qwen2_p128_l0_together.axmodel
        ...
        |-- llava_qwen2_p128_l9_together.axmodel
        |-- llava_qwen2_post.axmodel
        |-- model.embed_tokens.weight.bfloat16.bin
        |-- post_config.json
        `-- vision_cache

3 directories, 37 files

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

运行（CLI）

root@ax650:~# axllm run AXERA-TECH/FastVLM-1.5B-GPTQ-Int4/
[I][                            Init][ 138]: LLM init start
tokenizer_type = 3
 96% | ███████████████████████████████   |  30 /  31 [3.66s<3.78s, 8.20 count/s] init post axmodel ok,remain_cmm(10593 MB)
[I][                            Init][ 199]: max_token_len : 1024
[I][                            Init][ 202]: kv_cache_size : 256, kv_cache_num: 1024
[I][                            Init][ 205]: prefill_token_num : 128
[I][                            Init][ 209]: grp: 1, prefill_max_kv_cache_num : 1
[I][                            Init][ 209]: grp: 2, prefill_max_kv_cache_num : 128
[I][                            Init][ 209]: grp: 3, prefill_max_kv_cache_num : 256
[I][                            Init][ 209]: grp: 4, prefill_max_kv_cache_num : 512
[I][                            Init][ 209]: grp: 5, prefill_max_kv_cache_num : 640
[I][                            Init][ 214]: prefill_max_token_num : 640
[I][                            Init][  27]: LLaMaEmbedSelector use mmap
100% | ████████████████████████████████ |  31 /  31 [3.66s<3.66s, 8.47 count/s] embed_selector init ok
[W][                            Init][ 480]: classic vision size override: cfg=448x448 -> model=1024x1024 (from input shape)
[I][                            Init][ 666]: VisionModule init ok: type=FastVLM, tokens_per_block=256, embed_size=1536, out_dtype=fp32
[I][                     load_config][ 282]: load config:
{
    "enable_repetition_penalty": false,
    "enable_temperature": false,
    "enable_top_k_sampling": false,
    "enable_top_p_sampling": false,
    "penalty_window": 20,
    "repetition_penalty": 1.2,
    "temperature": 0.9,
    "top_k": 10,
    "top_p": 0.8
}

[I][                            Init][ 272]: LLM init ok
Type "q" to exit
Ctrl+c to stop current running
"reset" to reset kvcache
"dd" to remove last conversation.
"pp" to print history.
VLM enabled: after each prompt, input image path (empty = text-only). Use "video:<frames_dir>" for video.
----------------------------------------
prompt >> who are you
image >>
[I][                      SetKVCache][ 406]: prefill_grpid:2 kv_cache_num:128 precompute_len:0 input_num_token:22
[I][                      SetKVCache][ 408]: current prefill_max_token_num:640
[I][                      SetKVCache][ 409]: first run
[I][                             Run][ 457]: input token num : 22, prefill_split_num : 1
[I][                             Run][ 497]: prefill chunk p=0 history_len=0 grpid=1 kv_cache_num=0 input_tokens=22
[I][                             Run][ 519]: prefill indices shape: p=0 idx_elems=128 idx_rows=1 pos_rows=0
[I][                             Run][ 627]: ttft: 137.01 ms
I am an AI language model, I am here to help answer any questions you may have. How can I assist you today?

[N][                             Run][ 709]: hit eos,avg 14.77 token/s

[I][                      GetKVCache][ 380]: precompute_len:48, remaining:592
prompt >> describe the image
image >> ./AXERA-TECH/FastVLM-1.5B-GPTQ-Int4/image.png
[I][                EncodeForContent][ 994]: vision cache store: ./AXERA-TECH/FastVLM-1.5B-GPTQ-Int4/image.png
[I][                      SetKVCache][ 406]: prefill_grpid:4 kv_cache_num:512 precompute_len:48 input_num_token:271
[I][                      SetKVCache][ 408]: current prefill_max_token_num:512
[I][                             Run][ 457]: input token num : 271, prefill_split_num : 3
[I][                             Run][ 497]: prefill chunk p=0 history_len=48 grpid=2 kv_cache_num=128 input_tokens=128
[I][                             Run][ 519]: prefill indices shape: p=0 idx_elems=128 idx_rows=1 pos_rows=0
[I][                             Run][ 497]: prefill chunk p=1 history_len=176 grpid=3 kv_cache_num=256 input_tokens=128
[I][                             Run][ 519]: prefill indices shape: p=1 idx_elems=128 idx_rows=1 pos_rows=0
[I][                             Run][ 497]: prefill chunk p=2 history_len=304 grpid=4 kv_cache_num=512 input_tokens=15
[I][                             Run][ 519]: prefill indices shape: p=2 idx_elems=128 idx_rows=1 pos_rows=0
[I][                             Run][ 627]: ttft: 403.77 ms
The image depicts three astronauts standing in a forest, wearing full space suits with helmets. The scene is surreal and otherworldly, as the astronauts are dressed in space suits and are surrounded by a natural environment. The image is in black and white, which adds to the surreal and dreamlike quality of the scene. The astronauts appear to be exploring the forest, and the contrast between the natural environment and the space suits creates a striking and thought-provoking image.

[N][                             Run][ 709]: hit eos,avg 14.79 token/s

[I][                      GetKVCache][ 380]: precompute_len:412, remaining:228
prompt >> how many people in the image?
image >>
[I][                EncodeForContent][ 926]: vision cache hit (mem): ./AXERA-TECH/FastVLM-1.5B-GPTQ-Int4/image.png
[I][                      SetKVCache][ 406]: prefill_grpid:4 kv_cache_num:512 precompute_len:412 input_num_token:17
[I][                      SetKVCache][ 408]: current prefill_max_token_num:128
[I][                             Run][ 457]: input token num : 17, prefill_split_num : 1
[I][                             Run][ 497]: prefill chunk p=0 history_len=412 grpid=4 kv_cache_num=512 input_tokens=17
[I][                             Run][ 519]: prefill indices shape: p=0 idx_elems=128 idx_rows=1 pos_rows=0
[I][                             Run][ 627]: ttft: 168.52 ms
There are three people in the image.

[N][                             Run][ 709]: hit eos,avg 14.69 token/s

[I][                      GetKVCache][ 380]: precompute_len:437, remaining:203
prompt >> q

启动服务（OpenAI 兼容）

root@ax650:~# axllm serve AXERA-TECH/FastVLM-1.5B-GPTQ-Int4/
[I][                            Init][ 138]: LLM init start
tokenizer_type = 3
 96% | ███████████████████████████████   |  30 /  31 [2.72s<2.81s, 11.02 count/s] init post axmodel ok,remain_cmm(10593 MB)
[I][                            Init][ 199]: max_token_len : 1024
[I][                            Init][ 202]: kv_cache_size : 256, kv_cache_num: 1024
[I][                            Init][ 205]: prefill_token_num : 128
[I][                            Init][ 209]: grp: 1, prefill_max_kv_cache_num : 1
[I][                            Init][ 209]: grp: 2, prefill_max_kv_cache_num : 128
[I][                            Init][ 209]: grp: 3, prefill_max_kv_cache_num : 256
[I][                            Init][ 209]: grp: 4, prefill_max_kv_cache_num : 512
[I][                            Init][ 209]: grp: 5, prefill_max_kv_cache_num : 640
[I][                            Init][ 214]: prefill_max_token_num : 640
[I][                            Init][  27]: LLaMaEmbedSelector use mmap
100% | ████████████████████████████████ |  31 /  31 [2.72s<2.72s, 11.38 count/s] embed_selector init ok
[W][                            Init][ 480]: classic vision size override: cfg=448x448 -> model=1024x1024 (from input shape)
[I][                            Init][ 666]: VisionModule init ok: type=FastVLM, tokens_per_block=256, embed_size=1536, out_dtype=fp32
[I][                     load_config][ 282]: load config:
{
    "enable_repetition_penalty": false,
    "enable_temperature": false,
    "enable_top_k_sampling": false,
    "enable_top_p_sampling": false,
    "penalty_window": 20,
    "repetition_penalty": 1.2,
    "temperature": 0.9,
    "top_k": 10,
    "top_p": 0.8
}

[I][                            Init][ 272]: LLM init ok
Starting server on port 8000 with model 'AXERA-TECH/FastVLM-1.5B-GPTQ-Int4'...
OpenAI API Server starting on http://0.0.0.0:8000
Max concurrency: 1
Models: AXERA-TECH/FastVLM-1.5B-GPTQ-Int4

OpenAI 调用示例

from openai import OpenAI

API_URL = "http://127.0.0.1:8000/v1"
MODEL = "AXERA-TECH/FastVLM-1.5B-GPTQ-Int4"

messages = [
    {"role": "system", "content": [{"type": "text", "text": "you are a helpful assistant."}]},
    {"role": "user", "content": "hello"},
]

client = OpenAI(api_key="not-needed", base_url=API_URL)
completion = client.chat.completions.create(
    model=MODEL,
    messages=messages,
)

print(completion.choices[0].message.content)

OpenAI 流式调用示例

from openai import OpenAI

API_URL = "http://127.0.0.1:8000/v1"
MODEL = "AXERA-TECH/FastVLM-1.5B-GPTQ-Int4"

messages = [
    {"role": "system", "content": [{"type": "text", "text": "you are a helpful assistant."}]},
    {"role": "user", "content": "hello"},
]

client = OpenAI(api_key="not-needed", base_url=API_URL)
stream = client.chat.completions.create(
    model=MODEL,
    messages=messages,
    stream=True,
)

print("assistant:")
for ev in stream:
    delta = getattr(ev.choices[0], "delta", None)
    if delta and getattr(delta, "content", None):
        print(delta.content, end="", flush=True)
print("
")