Modal Examples

The project is organized into several modules that separate the web interface, agent logic, and graph structure:

• codelangchain.py: The main web application definition.
• agent.py: Contains the core LangChain agent definition.
• nodes.py: Defines the nodes within the LangGraph agent graph.
• edges.py: Defines the edges (control flow) within the LangGraph agent graph.
• retrieval.py: Contains the logic for retrieving documentation to include in the model's prompt.
• common.py: Contains Modal container image definitions and shared utilities.

The repository is organized into functional categories to help you learn Modal's capabilities:

• Guided Tour: Folders 01_getting_started/ through 14_clusters/ provide a progressive learning path through Modal's core concepts.
• Miscellaneous: The misc/ folder contains uncategorized examples.

Note: All examples are continuously tested for correctness against Python 3.11.

The misc/ directory contains a variety of examples demonstrating different ways to use Modal.

Note: Unlike other examples in the modal-examples repository, the examples in this directory are not continually monitored for correctness. They may become out of date or incorrect over time as the Modal API evolves.

You can verify that the Wikipedia dataset was successfully saved to the embedding-wikipedia volume by listing its contents using the Modal CLI.

To list the root of the volume:

modal volume ls embedding-wikipedia

To list the contents of the specific /wikipedia directory:

modal volume ls embedding-wikipedia /wikipedia

You can fine-tune the whisper-small model for improved automatic Hindi speech recognition using Modal. The provided configuration typically runs for approximately 3 hours and achieves a Word Error Rate (WER) of about 55-60. To improve performance and decrease the WER, you should increase the number of training epochs.

Benchmark performance using the Huggingface autoevaluate leaderboard for the mozilla-foundation/common_voice_11_0 dataset.

modal run -m train.train --num_train_epochs=10

Once the dataset is downloaded, use main.py to run the embedding job. This script utilizes Modal's built-in parallelization abstractions to process the dataset.

Note: The embedding process uses two separate Modal volumes: one for reading the source dataset and another for writing the resulting files for upload.

Run the embedding script using:

modal run main.py

To use these examples, first sign up for a free account at modal.com and follow the official setup instructions to install the modal package and configure your API key.

Once set up, you can run any example script using the modal run command. While you execute the command from your local machine, the code will communicate with Modal to spawn and run serverless containers in the cloud.

modal run 01_getting_started/hello_world.py

To ensure the fine-tuning program is functional, you can run an end-to-end check. This command executes a full train → serialize → save → load → predict cycle in less than 5 minutes.

modal run -m train.end_to_end_check

To run this example, you need to set up a Python virtual environment and install the modal package. You must also authenticate your local environment with Modal.

1. Create and activate a virtual environment:

   python3 -m venv venv
   source venv/bin/activate

2. Install the Modal client:

   pip3 install modal

3. Authenticate with Modal:

   modal token new

To manage storage costs, you can remove files or directories from a Modal Volume using the modal volume rm command.

Important: To remove a directory, you must include the --recursive flag.

Example to remove the /wikipedia directory from the embedding-wikipedia volume:

modal volume rm embedding-wikipedia /wikipedia --recursive

The download.py script downloads the Wikipedia dataset directly into a Modal Volume. This leverages Modal's high internet speeds for rapid data ingestion. To optimize download times, the script uses the num_proc keyword to parallelize downloads; setting this value between 4 and 10 can reduce download time by 30-40%.

Run the download script using:

modal run download.py