Mastering Agentic Coding and GPUs
summary
Shows an agentic coding platform that rewrote CUDA kernels for EFA networking to adapt Deepseek/DP-style kernels originally built for Infiniband, and reports a 20% performance improvement on AWS. Demonstrates automated build and rollout: automated Kubernetes deployment and Docker container management, AI agents that search code, generate and patch C++/CUDA, run clang/Perf/BFtrace analyses, and validate fixes against benchmarks while deploying Dynamo and GPU-accelerated inferencing.
AI-written summary · about
chapters (13)
- 0:00Start
- 0:14Introduction
- 3:04Agentic coding framework
- 7:05Building HPC agent stack
- 12:22Hardest coding problems solved
- 20:00Multi-LLM system and supervision
- 27:20Supervisor framework
- 37:00Deploying Kubernetes and endpoints
- 1:01:13Dynamo and disaggregated inferencing
- 1:12:13Self-hosted agent deployment
- 1:31:55Scientific process and feedback
- 1:41:11Human-in-the-loop evaluation
- 1:52:19Closing thoughts
related talks
18mYour Coding Agent Should Do AI System Engineering
06mThe Building Blocks of Agentic Al
21mUnlocking How To Efficiently, Flexibly, Manage and Schedule Seven AI Chips in Kubernetes
28mDeveloper Essentials for the New Era of AI
22mScaling agentic AI: From infrastructure to enterprise outcomes – Duncan Ng
34mGPU as a Service Over K8s: Drive Productivity and Increase Utilization
13mARK: GPU-driven Code Execution for Distributed Deep Learning
57mDeep Learning with C++