Learn in public, with reproducible questions
Useful engineering discussions begin with a small reproducer and evidence. Use official project channels linked from these sources.
Serving engines
Before filing an issue, record engine/model versions, hardware, flags, a synthetic reproducer, and expected versus observed behavior.
Documentation · vLLM · Optimization and tuning ↗
Documentation · SGLang · Serving documentation ↗
Kernels and communication
Share a short trace, tensor shapes, dtype, topology, and measurements. Remove private prompts and credentials.
Tutorials · Triton · GPU programming tutorials ↗
Documentation · NVIDIA · NCCL collectives ↗
Study groups
Explain one paper diagram and one limitation, then run a small comparison together.
Course + videos · Stanford · CS336 Language Modeling from Scratch (2025) ↗
Course + videos · UC Berkeley · CS285 Deep Reinforcement Learning ↗