doc/vllm_ascend_for_openpangu_vl_7b.md

openPangu-VL-7B 在vllm-ascend部署指导文档

部署环境要求

Atlas 800T A2(64GB) 1、2、4、8卡均可部署openPangu-VL-7B

镜像构建和启动

选用vllm-ascend社区镜像v0.9.1

拉取方式如下：

docker pull quay.io/ascend/vllm-ascend:v0.9.1

以下操作需在每个节点都执行。 启动镜像。

# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:v0.9.1  # Use correct image id
export NAME=vllm-ascend  # Custom docker name

# Run the container using the defined variables
# Note if you are running bridge network with docker, Please expose available ports for multiple nodes communication in advance
# To prevent device interference from other docker containers, add the argument "--privileged"
docker run --rm \
--name $NAME \
--network host \
--ipc=host \
--device /dev/davinci0 \
--device /dev/davinci1 \
--device /dev/davinci2 \
--device /dev/davinci3 \
--device /dev/davinci4 \
--device /dev/davinci5 \
--device /dev/davinci6 \
--device /dev/davinci7 \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /mnt/sfs_turbo/.cache:/root/.cache \
-it $IMAGE bash

需要保证模型权重和本项目代码可在容器中访问。如果未进入容器，需以root用户进入容器。

docker exec -itu root $NAME /bin/bash

PD混部推理

示例启动脚本：LOAD_CKPT_DIR=xxx bash examples/start_serving_openpangu_vl_7b.sh。该启动脚本为8卡推理（变量TENSOR_PARALLEL_SIZE_LOCAL=8）。拉起服务后，可向首节点（主节点）发送请求。

发请求测试

服务启动后，可发送测试请求。推荐使用示例中的system prompt。

推理示例：图片+文字

import json
import base64
import os
import requests
import json

def encode_image_to_base64(img_path, img_name):
    #load image to base64
    try:
        with open(os.path.join(img_path, img_name), 'rb') as img_file:
            img_data = img_file.read()
        base64_str = base64.b64encode(img_data).decode('utf-8')
        return base64_str
    except Exception as e:
        print(f"image load failed: {e}")
        return None

base64_image = encode_image_to_base64("/image_path", "image_name.jpg")


payload_image_example = json.dumps({
  "messages": [
      {
            "role": "system",
            "content": [
                {"type": "text", "text": "你是华为公司开发的多模态大模型，名字是openPangu-VL-7B。你能够处理文本和视觉模态的输入，并给出文本输出。"},     
            ]
      },
      {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": f"data:image/jpg;base64,{base64_image}"}},
                {"type": "text", "text": "Please describe this picture."}, 
            ]
      }
  ],    
  "model": "pangu_vl",
  "max_tokens": 500,
  "temperature": 1.0,
  "stream": False,
})


url = "http://127.0.0.1:8000/v1/chat/completions"
headers = {
  'Content-Type': 'application/json'
}

response_image_example = requests.request("POST", url, headers=headers, data=payload_image_example)
print(f"the response of image example is {response_image_example.text}")

推理示例：视频+文字

import json
import base64
import os
import requests
import json

def encode_video_to_base64(video_path, video_name):
    #load video to base64
    try:
        with open(os.path.join(video_path, video_name), 'rb') as video_file:
            video_data = video_file.read()
        base64_str = base64.b64encode(video_data).decode('utf-8')
        return base64_str
    except Exception as e:
        print(f"video load failed: {e}")
        return None

base64_video = encode_video_to_base64("/video_path", "video_name.mp4")

payload_video_example = json.dumps({
  "messages": [
      {
            "role": "system",
            "content": [
                {"type": "text", "text": "你是华为公司开发的多模态大模型，名字是openPangu-VL-7B。你能够处理文本和视觉模态的输入，并给出文本输出。"},     
            ]
      },
      {
            "role": "user",
            "content": [
                {"type": "video_url", "video_url": {"url": f"data:video/mp4;base64,{base64_video}"}},
                {"type": "text", "text": "Please describe this video."},  
            ]
      }
  ],    
  "model": "pangu_vl",
  "max_tokens": 500,
  "temperature": 1.0,
  "stream": False,
})


url = "http://127.0.0.1:8000/v1/chat/completions"
headers = {
  'Content-Type': 'application/json'
}

response_video_example = requests.request("POST", url, headers=headers, data=payload_video_example)
print(f"the response of video example is {response_video_example.text}")

128k 视频长序列推理

在/preprocessor_config.json中添加字段,输入视频会被抽取为768帧

"num_frames": 768,
"sample_fps": -1.0

启动脚本(/inference/vllm_ascend/examples/start_serving_openpangu_vl_7b.sh)内设置参数：

export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
MAX_MODEL_LEN=128000
MAX_NUM_BATCHED_TOKENS=100000
GPU_MEMORY_UTILIZATION=0.7

--no-enable-chunked-prefill \
--no-enable-prefix-caching \

Int8推理

ModelSlim量化

openPangu-VL-7B模型支持使用开源量化框架ModelSlim,参考[ModelSlim_openPangu-VL-7B]，当前模型支持W8A8权重激活量化。

openPangu-VL-7B W8A8 动态量化

export QUANT_PATH=your_quant_save_dir
export MODEL_PATH=your_model_ckpt_dir
export CALI_DATASET=your_cali_dataset_dir
python quant_pangu_vl.py \
--model_path $MODEL_PATH --calib_images $CALI_DATASET \
--save_directory $QUANT_PATH --w_bit 8 --a_bit 8 --device_type npu \
--trust_remote_code True --anti_method m2 --act_method 3 --is_dynamic True

相较于BF16模型，int8量化模型的config.json增加以下字段：

"quantize": "w8a8_dynamic",

ModelSlim量化脚本生成量化模型后会自动追加上述字段到config.json中。

Int8推理

相较于BF16模型推理，int8量化模型推理可使用同样的启动脚本，仅需：

• 减少节点数、卡数；
• 修改模型Checkpoint路径。