MLOps vs LLMOps
summary
Compares traditional MLOps pipelines (code, model registry, baseline datasets, and embedding lifecycle) with LLM-first architectures that use vendor pre-trained LLMs and treat embeddings and prompt lifecycle as production artifacts. Presents a sentiment-analysis pipeline with versioned models plus an internal RAG system backed by a vector DB that records Q/A feedback for RLHF and to measure cost/latency tradeoffs. Highlights the need for monitoring data drift, prompt/embedding combinations and embedding versioning, and notes higher compute needs and a larger attack surface for LLMOps.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:20DevOps methods
- 4:41ML lifecycle
- 9:04Model decay monitoring
- 12:17Versioning models/data
- 19:13LLM introduction
- 23:11LLMOps lifecycle with RAG
- 29:45MLOps vs LLMOps
- 34:04LLMOps resources & monitoring
- 35:57Conclusion & Q&A
related talks
49mLLMops: what's beyond ChatGPT and how you can deploy it
27mWHAT DOES MLOps LOOK LIKE FOR LARGE LANGUAGE MODELS (LLMs) IN PRODUCTION?
29mCustomizing and Evaluating LLMs, an Ops Perspective
30mLLM-Based Recommendation Systems
15mOpenAssistant's Opensource LLMs Saga
24mLLM Inference Optimization From Simple To Sophisticated
28mNavigating LLM deployment: Tips, tricks, and techniques
52mAnatomy of an LLM