[Hands-on Workshop] vLLM Reality Check: Understand, Deploy, Optimize, Automate

​What the Workshop Will be About:

​This workshop will provide a deep dive into vLLM, a high-performance inference engine for large language models.

​Participants will explore how vLLM works under the hood, how it optimizes model execution, and how to effectively deploy and manage it in production environments.

​Key Points to Cover:

• ​vLLM Internals - Understand the request lifecycle. How vLLM manages model families, and how GPU execution is orchestrated for optimal performance.

• ​Transformer Architecture & vLLM Optimizations - Revisit transformer architecture foundations and learn how vLLM leverages techniques like continuous batching, PagedAttention to accelerate inference.

• ​Advanced Deployment Strategies - Explore best practices for deploying vLLM across different environments, starting from single-GPU setups.

• ​Automated management of inference engines - Discuss what are shortcomings of self serving, why inference providers exist, and test a deployment flow in Cast.ai

​Expected Outcomes:

• ​Unpack the vLLM deployment struggle

• ​Benchmark performance gaps between manual and optimized configurations

• ​Discover GPU optimization challenges - availability, selection, and scaling issues

• ​See automated deployment solve it - same benchmark, better results

• ​You'll be the one running the commands, hitting the walls, and discovering why leading teams automate their AI infrastructure with AI Enabler instead of building it from scratch.

​Who it’s for:

​AI/ML Engineers, MLOps, LLMOps, DevOps Engineers, and Platform Engineers running or planning to run models in production.

​Pre-Workshop Requirements:

​A laptop & basic knowledge of Kubernetes and LLMs

​Meet Your Conductor:

​Igor Šušić, a talented Staff Machine Learning Engineer will be driving the masterclass.

​Food, drinks and good vibes will be provided during the workshop.