Large Model User Guide

Introduction

This document provides an overview of how to obtain and use large models, including models for text-to-image, image-to-image, speech-to-text, text-to-speech, and chat.

List of supported large models:

Download Methods

Currently, two download methods are provided: cloud storage download and HuggingFace download.

Cloud Storage Download

Download Images (Baidu Netdisk) Extraction code:dshn

Download Images (MEGA)

From the List of supported large models above, locate the required model and download it from the cloud storage.
For example, for the lcm-lora-sdv1-5-maixcam2 model, download a file similar to lcm-lora-sdv1-5-maixcam2-202601051759.zip, The suffix 202601051759 indicates the model packaging time.

HuggingFace Download

注：

1. Downloading from HuggingFace requires a stable network environment. Poor connectivity may result in interrupted downloads
2. The following methods can also be executed directly in the terminal of the target platform (e.g., MaixCAM2).

1. Download via Command Line

   Install the download tool:

   pip install huggingface_hub
   

   Set the download endpoint. The default is https://huggingface.co
   For users in China, it is recommended to use https://hf-mirror.com

   # Linux/MacOS
   export HF_ENDPOINT=https://hf-mirror.com # or 'https://huggingface.co'
   
   # Windows
   ## cmd
   set HF_ENDPOINT=https://hf-mirror.com
   ## PowerShell
   $env:HF_ENDPOINT = "https://hf-mirror.com"
   

   Example for downloading the lcm-lora-sdv1-5-maixcam2 model.
   To download another model, replace lcm-lora-sdv1-5-maixcam2 with the desired model name.

   # Download model (new)
   hf download sipeed/lcm-lora-sdv1-5-maixcam2 --local-dir /root/models
   
   # Download model (legacy)
   huggingface-cli download sipeed/lcm-lora-sdv1-5-maixcam2 --local-dir /root/models
   

2. Download Using Python

   Install the huggingface_hub package:

pip install huggingface_hub

Example for downloading the lcm-lora-sdv1-5-maixcam2 model. To download other models, replace the model_name variable accordingly.

# This scripy is used to install models from huggingface
# Only support MaixCAM2 platform

import os
os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com' # or 'https://huggingface.co'

from huggingface_hub import snapshot_download
from huggingface_hub.utils import tqdm

CAPTURE_PROGRESS=False
class Tqdm(tqdm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        if CAPTURE_PROGRESS:
            print(f"[INIT] {self.desc} | {self.n}/{self.total} ({self.n/self.total*100:.2f}%)")
        else:
            print('')

    def update(self, n=1):
        super().update(n)
        if CAPTURE_PROGRESS:
            print(f"[UPDATE] {self.desc} | {self.n}/{self.total} ({self.n/self.total*100:.2f}%)")
        else:
            print('')

    def close(self):
        super().close()
        if CAPTURE_PROGRESS:
            print(f"[CLOSE] {self.desc} | {self.n}/{self.total} ({self.n/self.total*100:.2f}%)")
        else:
            print('')

model_name = 'lcm-lora-sdv1-5-maixcam2'
repo_id = f'sipeed/{model_name}'
local_dir = f'/root/models/{model_name}'
snapshot_download(
    repo_id=repo_id,
    local_dir=local_dir,
    # allow_patterns="*.py",
    tqdm_class=Tqdm,
)

Uploading Models to the Board

After downloading, you can upload the model to the board using the scp command. It is recommended to upload models to the /root/models directory.

Using lcm-lora-sdv1-5-maixcam2 as an example:

1. Confirm the Files to Upload

   The directory structure of the lcm-lora-sdv1-5-maixcam2 model is:

   lcm-lora-sdv1-5-maixcam2
   ├── lcm-lora-sdv1-5-maixcam2					# The actual model files to upload
   ├── README.md
   ├── README_ZH.md
   └── launcher.py
   

   The folder lcm-lora-sdv1-5-maixcam2 under the parent directory is the actual model directory that must be uploaded. Make sure to upload the entire model directory.

2. Upload the Model to the Development Board

   Example of uploading the lcm-lora-sdv1-5-maixcam2 model to the board using scp(assuming the board IP address is 192.168.10.100):

   Notes:

   1. Uploading models through the USB network connection is usually faster. See “Preparing to Connect the Computer and Device” in MaixCAM2 Quick Start to find its IP address.

   scp -r lcm-lora-sdv1-5-maixcam2/lcm-lora-sdv1-5-maixcam2 root@192.168.10.100:/root/models
   

Using the Models

For instructions on how to use each model, please refer to the corresponding documentation, for example:
Qwen Large Language Model
InternVL Vision-Language Model
SmolVLM Vision-Language Model