Skip to content

demo.md

Latest commit

 

History

History
95 lines (68 loc) · 3.27 KB

demo.md

File metadata and controls

95 lines (68 loc) · 3.27 KB

Demo

This demo shows how to use the Union MCP server to run compute- and io-intensive tools on a Flyte cluster.

Setup

Take a look at the mcp configuration guide for more detailed instructions on how to configure your MCP client to use the Union MCP server.

Cursor

Add the following to your ~/.cursor/mcp.json file:

{
  "mcpServers": {
    "union-mcp-v2": {
      "url": "https://mcp-v2.apps.demo.hosted.unionai.cloud/sdk/mcp"
      "headers": {
        "Authorization": "Bearer <secret-value>"
      }
    }
  }
}

Claude Code

Add the following to your ~/.claude.json file:

{
  "mcpServers": {
    "union-mcp-v2": {
      "url": "https://mcp-v2.apps.demo.hosted.unionai.cloud/sdk/mcp"
      "headers": {
        "Authorization": "Bearer <secret-value>"
      }
    }
  }
}

Claude Desktop

Configure the claude_desktop_config.json configuration file located in:

  • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
  • Windows: %APPDATA%\Claude\claude_desktop_config.json

Make sure that npx is installed and available in your $PATH.

{
  "mcpServers": {
    "union-mcp-v2": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://mcp-v2.apps.demo.hosted.unionai.cloud/sdk/mcp",
        "--header",
        "Authorization: Bearer ${AUTH_TOKEN}"
      ],
      "env": {
        "AUTH_TOKEN": "<secret-value>"
      }
    }
  }
}

Prompts

Toy example

Create a flyte script that fans out tasks to compute the square of the numbers from 1 to 100_000, where each task handles 1_000 numbers, then sums the squares.

Visualization

Create and run a flyte script that downloads the dataset at https://github.com/plotly/datasets/blob/master/timeseries.csv and creates a visualization in plotly and run it remotely. The flyte script should use flyte.report to render a beautiful visualization.

Model training

Run a flyte script that performs hyperparameter optimization that uses flyte to parallelize the training runs for training a random forest model on the penguins data. Assess f1 score as the evaluation metric, and visualize the results using flyte.report. Make sure the report style is beautiful.

PDF transcription

Create a flyte script that grabs the 5 most recently published articles from Arxiv and transcribes them using microsoft/VibeVoice-1.5B using the huggingface transformers library. Use a driver-worker pattern where the driver is a CPU environment and the worker is a GPU environment with one T4 GPU. Save the transcriptions to a json file using flyte.io.File, and use flyte.report make a pretty visualization of the transcriptions, including a preview of text the contents of documents the audio files embedded in the html report.

GPU Batch inference

Run a flyte script that embeds the "review" column of the "scikit-learn/imdb" huggingface dataset using the "answerdotai/ModernBERT-base" model on a T4 GPU. Use a driver-worker pattern where the driver is a CPU environment and the worker is a GPU environment. Save the embeddings to a json file using flyte.io.File, and use flyte.report make a pretty visualization of the embeddings, including a preview of text the contents of the first five documents and the distribution of their embeddings.