This repository contains the accompanying materials for talks given as part of the AMD ROCm™ Webinar Series presented by Fluid Numerics for Advanced Micro Devices, Inc. (AMD).
The contents of these repositories are considered up-to-date for the date the talk is given. No materials should be expected to be maintained, updated versions, etc. past these dates. Many notebooks/scripts/etc will likely still work for recent versions of software with a simple version change in the appropriate file, but this is not guranteed.
E.g., getting-started-with-llama-3/install.sh will never be updated to install ROCm 6.2 or beyond, since ROCm 6.1 was the latest version at the day of presentation.
Topics Covered:
- Fundamentals of transformer models
- How to install torch and transformers libraries and their dependencies on AMD Instinct and Radeon GPUs with ROCm 6
- Running Llama-3-8B and serving on AMD MI210 and AMD W7800 GPUs
- Discussion on rudimentary techniques such as temperature adjustment and greedy selection
Accompanying material can be found under getting-started-with-llama-3.
Slides can be found here. Presented 11 July 2024.
Topics Covered:
- When is fine-tuning necessary?
- Dataset selection for fine-tuning
- The role of quantization in training and inference
- Techniques and tools such as LoRA, QLoRA, PEFT, and FSDP
- Running a fine-tuned model on AMD Radeon GPUs
Accompanying material can be found under fine-tuning-llama-3.
Slides can be found here. Presented 15 October 2024.