接下來會透過 edX - Large Language Modes: Application through Production 課程提到的 LLMOps Notebook 來探討。

LLMOps Notebook

流程如下：

Develop an LLM pipeline

• Load data.
• Build an LLM pipeline.
• Test applying the pipeline to data, and log queries and results to MLflow Tracking.
• Log the pipeline to the MLflow Tracking server as an MLflow Model.

Test the LLM pipeline

• Register the pipeline to the MLflow Model Registry.
• Test the pipeline on sample data.
• Promote the registered model (pipeline) to production.

Create a production workflow for batch inference

• Load the latest production LLM pipeline from the Model Registry.
• Apply the pipeline to an Apache Spark DataFrame.
• Append the results to a Delta Lake table.

下一篇會說明 Notebook 內容與在 Databricks UI 的驗證。

Reference:

• https://www.edx.org/course/large-language-models-application-through-production
• https://github.com/databricks-academy/large-language-models