Merge remote-tracking branch 'upstream/main'

Author: dchourasia

examples/kfto-feast/README.md

@@ -0,0 +1,112 @@
+# 🚀 Quickstart: Fine‑Tuning Granite Models with FSDP/DeepSpeed & LoRA/QLoRA Using a Simple FEAST Store example
+
+This notebook guides you through the process of fine-tuning **Large Language Models** using **Feast**, **Kubeflow-Training**, and modern optimization strategies like **FSDP**, **DeepSpeed**, and **LoRA** to boost training performance and efficiency.
+
+In particular, this example demonstrates:
+1. How to implement **Fully Sharded Data Parallel (FSDP)** and **DeepSpeed** to distribute training across multiple GPUs, enhancing scalability and speed.
+2. How to apply **Low-Rank Adaptation (LoRA)** or **Quantized Low-Rank Adaptation (QLoRA)** via the [PEFT library](https://github.com/huggingface/peft) for parameter-efficient fine-tuning, reducing computational and memory overhead.
+3. How to retrieve and manage **training features using Feast**, enabling consistent, scalable, and reproducible ML pipelines.
+
+---
+
+## 🍽️ What is Feast and How Are We Using It?
+
+[Feast (Feature Store)](https://github.com/feast-dev/feast) is a powerful operational data system for machine learning that helps manage, store, and serve features consistently during training and inference. In this workflow, **Feast acts as the centralized source of truth for model features**, decoupling feature engineering from model training.
+
+Specifically, we use Feast to:
+
+- **Define and register feature definitions** for training data using a standardized interface.
+- **Ingest and materialize features** from upstream sources (e.g., batch files, data warehouses).
+- **Fetch training features** as PyTorch-friendly tensors, ensuring consistency across training and production.
+- **Version control feature sets** to improve reproducibility and traceability in experiments.
+
+By integrating Feast into the fine-tuning pipeline, we ensure that the training process is not only scalable but also **robust, modular, and production-ready**.
+
+---
+
+## 💡 Why Use FSDP, DeepSpeed, LoRA, and Feast for Fine-Tuning?
+
+- **Efficient Distributed Training:** Utilize FSDP or DeepSpeed to handle large models by distributing the training process, enabling faster and more scalable training sessions.
+- **Parameter Efficiency with LoRA/QLoRA:** Implement LoRA to fine-tune models by updating only a subset of parameters, LoRA’s low‑rank adapters let you adapt a full LLM by training a tiny fraction of parameters, saving compute and storage by ~90% and speeding up training ; QLoRA extends the LoRA technique by loading the model in 4‑bit quantization, freezing the quantized weights, and updating only the adapters to support massive models on limited‑memory GPUs.
+- **Feature Management with Feast:** Fetch well-defined, version-controlled features seamlessly into your pipeline, boosting reproducibility and easing data integration.
+- **Flexible Configuration Management:** Store DeepSpeed and LoRA settings in separate YAML files, allowing for easy modifications and experimentation without altering the core training script.
+- **Mixed-Precision Training:** Leverage automatic mixed precision (AMP) to accelerate training and reduce memory usage by combining different numerical precisions.
+- **Model Saving and Uploading:** Save the fine-tuned model and tokenizer locally and upload them to an S3 bucket for persistent storage and easy deployment.
+
+---
+
+## Requirements
+
+* An OpenShift cluster with OpenShift AI (RHOAI) 2.17+ installed:
+  * The `dashboard`, `trainingoperator` and `workbenches` components enabled
+* Sufficient worker nodes for your configuration(s) with NVIDIA GPUs (Ampere-based or newer recommended)
+  * If using PEFT LoRA/QLoRA techniques, then can use NVIDIA GPUs (G4dn)
+* AWS S3 storage available
+
+---
+
+
+By following this notebook, you'll gain hands-on experience in setting up a **feature-rich, efficient, and scalable** fine-tuning pipeline for **Granite language models**, leveraging tooling across model training and feature engineering.
+
+
+## Setup
+
+* Access the OpenShift AI dashboard, for example from the top navigation bar menu:
+
+![](./docs/01.png)
+
+* Log in, then go to _Data Science Projects_ and create a project:
+
+![](./docs/02.png)
+
+* Once the project is created, click on _Create a workbench_:
+  
+![](./docs/03.png)
+
+* Then create a workbench with the following settings:
+  
+    * Select the `PyTorch` (or the `ROCm-PyTorch`) notebook image:
+
+    ![](./docs/04a.png)
+
+    * Select the _Small_ container size and a sufficient persistent storage volume. 
+    * In _'Environment variables'_ section, use variable type as _Secret_ and provide key/value pair to store _HF-TOKEN_ as a kubernetes secret :
+  
+    ![](./docs/04b.png)
+
+    * Click on _Create connection_ to create a workbench connection to your S3 compatible storage bucket:
+
+        * Select option : _S3 compatible object storage - v1_
+
+
+        ![](./docs/04ci.png)
+
+        * Fill all the needed fields, also specify _Bucket_ value (it is used in the workbench), then confirm:
+
+
+        ![](./docs/04cii.png)
+
+        > [!NOTE]
+        >
+        > * Adding an accelerator is only needed to test the fine-tuned model from within the workbench so you can spare an accelerator if needed.
+        > * Keep the default 20GB workbench storage, it is enough to run the inference from within the workbench.
+        > * If you use different connection name than _s3-data-connection_ then you need to adjust the _aws_connection_name_ properly in notebook to refer to this new name.
+
+
+    * Review the configuration and click _Create workbench_:
+ 
+    ![](./docs/04d.png)
+
+* From "Workbenches" page, click on _Open_ when the workbench you've just created becomes ready:
+
+![](./docs/05.png)
+
+* From the workbench, clone this repository, i.e., `https://github.com/opendatahub-io/distributed-workloads.git`
+
+* Navigate to the `distributed-workloads/examples/kfto-feast` directory and open the `kfto_feast` notebook
+
+
+
+You can now proceed with the instructions from the notebook. Enjoy!
+
+