Keynote: DBRX: Training Modern LLMs on Databricks
summary
Demonstrates concrete LLM training pipelines on Databricks that combine Delta Lake for curated training data, MLflow for experiment and model tracking, and distributed runtimes using ZeRO/FSDP plus LoRA for parameter‑efficient fine‑tuning. Shows end‑to‑end runs scaling from single‑node to multi‑node GPU clusters, publishes measured throughput and cost trends, provides a reproducible notebook and model artifact on Databricks, and reports whether the produced checkpoints and inference optimizations meet expected accuracy and latency targets.
AI-written summary · about
chapters (4)
- 0:00DBRX and Databricks intro
- 4:58DBRX model capabilities
- 7:58Customers enabled with DBRX
- 11:49DBRX future vision
related talks
29mML in Production on Databricks
49mGenAI for Business on the Modern Data Stack
20mKeynote Review
28mDatabricks' vLLM Optimization for Cost-Effective LLM Inference
08mKeynote: Plug in and Scale: Serving LLM Models on Kubernetes Ma...
33mYour best Bet: Effortless MLOps with Python Models in
43mHow We Built DBRX - An Open High-quality LLM
51mHow to get started with MLOps in Azure Databricks