Azure OpenAI Batch Accelerator

Disclaimer:

This is a reference implementation of the Azure OpenAI Batch API designed to be extended for different use cases.
This code is NOT intended for production use but instead as a starting point/reference implenentation of the Azure OpenAI (AOAI) Batch API. The code here is provided AS IS, you assume all responsibility (e.g., charges) from running this code. Testing in your environment should be done before running large use cases. Lastly, this is a work in progress and will be updated frequently. Please check back regularly for updates.

Installation & Setup

Environment:

1. Python 3.11 (or higher)
2. Pip
3. An Azure Data Lake Storage (v2) account
4. An Azure OpenAI deployment

The following pip packages are required:

1. azure-storage-file-datalake
   
2. openai
3. tiktoken
4. requests
5. token-count
6. asyncio

In addition to this, it is recommended to install these dependencies in a virtual environment to avoid conflicts (e.g., .venv)

Connecting AOAI to Azure Storage

The `Storage Blob Data Contributer` role must be given to the AOAI service's Managed Identity to allow AOAI to access the data in the Azure Storage Account.

Configuration:

There are three configuration files required to use this accelerator:

1. AOAI_config.json - This file contains the settings for AOAI.
2. storage_config.json - This file contains the settings for the Azure Data Lake Storage Account which will hold the input/output of the job.
3. app_config.json - This file contains the application configuration settings.
4. APP_CONFIG in runBatch.py - This variable should be set to point to the app_config.json file which defines the app settings. Alternatively, this value can be set as an environment variable in the underlying OS. This will support command line parameter-based input in the future.

Reference templates of these files have been provided in the templates directory where <> denote settings that must be filled in. Other important settings are:

1. aoai_api_version - This must be set to 2024-07-01-preview as that's the only API version which supports the Batch API at this time. In the future, different versions can be set here.
2. batch_job_endpoint - This must be set to /chat/completions.
3. batch_size - This controls the 'micro batch' size which is the number of files that will be sent to the batch service in paralle. It is set to a recommended value of 10 but can be changed based on the requirements/file sizes being sent to the batch service.
4. download_to_local - This controls if the files should be downloaded to local to count the number of tokens in a file. Currently this should be set to the default value of false but may be used in future versions.
5. input_directory/filesystem - This is the directory and filesystem the code will check for input files, respectively. The default directory setting of / assumes no directories in the input filesystem. The current implementation is not recursive; if input files are stored in a directory in the input filesystem/container then it should be specified here.
6. output_directory/filesystem - This is the directory and filesystem the code will write output files, respectively. The default directory setting of / assumes no directories in the ouput filesystem.
7. error_directory/filesystem - This is the directory and filesystem the code will write error files, respectively. The default directory setting of / assumes no directories in the error filesystem.
8. continuous_mode - This setting controls how the code is run. If set to true, it will continuously check the input directory for files every 60 seconds, taking a snapshot of the files and kicking off a series of batch jobs to process until all files are processed. To stop, press ctrl+c. If set to false it will only run when executed.

Using the accelerator

1. Input: Upload formatted batch files to the input location specified in the storage_config.json configuration file. Once all files are uploaded, start the runBatch.py in the code directoy. When run, the code will run continuously or once, depending on the continuous_mode setting described above.
2. Output: The code will create a directory in the processed_filesystem_system_name location in storage_config.json configuration file for each file processed along with a timestamp of when the file was processed. The raw input file will also be moved to the processed directory. In addition, if there are any errors, they will be put in the error_filesystem_system_name location, with a timestamp.
3. Metadata: The output creates a metadata file for each input file which contains mapping information which may be useful for automated processing of results.
4. Cleanup: After processing is complete, the code will automatically process and clean up all files in the input directory, locally downloaded files, and all uploaded files to the AOAI Batch Service.

Authors & Contact

Please send any questions/comments to:

Amit Mukherjee amimukherjee@microsoft.com
DJ Dean dade@microsoft.com