D-Robotics Official Website
D-Robotics Official Website
English

Development Process

The overall development steps of Open D-Robotics-LLM mainly include environment deployment, large model quantization, and end-side deployment.

Environment Deployment

In this stage, you need to complete the correct installation and deployment of the development environment as required, to proceed with the subsequent steps. For details, please refer to the Environment Deployment section.

Large Model Quantization

In this stage, you need to download the required large language model from Hugging Face. We will convert this model into a format that can be deployed on the D-Robotics platform for subsequent inference. The steps include:

  1. Obtain the original model. Supported model types include:

    • Open-source Hugging Face format models: DeepSeek-R1-Distill-Qwen-1.5B、DeepSeek-R1-Distill-Qwen-7B、InternLM2-1.8B、Qwen2.5-1.5B、Qwen2.5-7B、Qwen2.5-1.5B-Instruct、Qwen2.5-7B-Instruct、Qwen2.5-Omni-3B.

    • Custom-trained large language models, which must have the same structure as the Hugging Face format models listed above.

  2. Model quantization (you can also skip this step by directly using the converted models we provide):

    Use the oellm_build command-line tool for conversion and compilation, building model that can be deployed on the D-Robotics platform. During the build process, you can set quantization parameters according to your actual needs. The generated *.hbm file can be used for subsequent board-side deployment.

Board-side Deployment

In this stage, you can deploy and run inference with the quantized models obtained in the previous stage. The steps include:

  1. Cross-compilation toolchain configuration: correctly extract the cross-compilation toolchain package and configure the path.

  2. Use the provided build scripts to generate executable files for board-side deployment.

  3. Board-side inference: after preparing all models, executables, and dependencies for board-side deployment, run the executable on the board device. Once completed, you can perform inference and dialogue tests.

  4. Resource release: after completing the workflow, destroy the inference task interface and release occupied resources.

Environment deployment, model quantization, and board-side deployment together form the complete D-Robotics-LLM development process, ensuring that large language models can be successfully converted, optimized, and efficiently deployed on the D-Robotics platform.