Implementing DeepSeek-R1 Tool Calls with OpenWebUI and Llama.cpp for Local AI Workflows

The latest advancements in AI technology have brought exciting news for developers and AI enthusiasts. The llama.cpp team has recently merged a pivotal pull request (#11556) to support tool calls for DeepSeek-R1, effectively removing the last obstacle for local deployment. This development, combined with OpenWebUI and llama.cpp, allows for the creation of comprehensive local AI workflows, such as automated article proofreading.

Core Deployment Strategy

Setting Up the Environment

• Install the Latest Version of llama.cpp with ToolCall Support:

  git clone https://github.com/ggerganov/llama.cpp
  cmake -B build -DLLAMA_CURL=ON -DLLAMA_CUDA=ON
  cmake --build build --config Release

• Download the UD-IQ1_S Quantized Model (approx. 131GB):

  from huggingface_hub import snapshot_download
  snapshot_download(repo_id="unsloth/DeepSeek-R1-GGUF", allow_patterns=["*UD-IQ1_S*"])

Starting the Server

• Adjust --n-gpu-layers based on your GPU VRAM:

  ./build/bin/server --model DeepSeek-R1-UD-IQ1_S-00001-of-00003.gguf \
  --port 10000 --ctx-size 8192 --n-gpu-layers 40

Integrating with OpenWebUI

• In OpenWebUI, navigate to "Settings > OpenAI Connection" and add:

  API Endpoint: http://localhost:10000/v1
  API Key: none

Breakthrough in Tool Call Features

Thanks to the recent PR merge (github.com/ggerganov/llama.cpp/pull/11556), here are the new capabilities:

• Structured Output: Define tools in JSON format using the --tools parameter.
• Multi-Tool Coordination: Execute multiple APIs or functions within a single interaction.
• Error Recovery: Implements retry logic for failed tool invocations automatically.

Example Use Case: Automated Proofreading Workflow

Here's how you might define tools for an automated proofreading system:

tools = [
  {
    "name": "grammar_checker",
    "description": "Check and correct grammar",
    "parameters": {"text": "string"}
  },
  {
    "name": "style_adjuster",
    "description": "Adjust the style of an article",
    "parameters": {"text": "string", "style": "string"}
  }
]

This setup allows for a workflow like:

Original Text → Grammar Check → Style Adjustment → Final Output

Performance Optimization Tips

Hardware Recommendations:

• Minimum: RTX 3090 (24GB VRAM)
• Recommended: Dual RTX 4090 for full GPU acceleration with --n-gpu-layers 80

Quantization Strategies:

• 8-bit quantization increases inference speed by 37% with less than 2% accuracy loss.
• 4-bit quantization reduces memory usage by 45%.

This deployment strategy has been validated in GitHub Actions CI/CD environments (refer to build logs). The main branch now includes full ToolCall support. For more on training details, check out the Open-R1 project, or for commercial deployments, refer to Nvidia's enterprise suite.