Systems learning hub
Production AI Systems
Production AI engineering turns a validated model into a reliable service. The work spans distributed training, attention and memory efficiency, model compression, batching, capacity, serving, observability, release gates, and rollback under real latency, cost, and reliability constraints.
Best for: Advanced learners and engineers operating model training or inference systems
Recommended learning order
- 1Attention Variants & NormsGrouped-Query Attention (GQA), Multi-Query (MQA), RoPE embeddings, and RMSNorm topology.Open lesson
- 2Distributed TrainingParallelism paradigms (DDP, TP, PP, FSDP) and scaling laws.Open lesson
- 3FlashAttention & GPU ArchitectureGPU HBM vs SRAM bandwidth, online softmax tiling, IO-awareness, and Triton CUDA ops.Open lesson
- 4Quantization & LoRALow-Rank Adaptation (ΔW = B·A), rank selection, FP8/INT4 uniform quantization, and scale-zero point math.Open lesson
- 5Mixture of Experts (MoE)Sparse routing mechanisms, gating networks, Top-K dispatch, and auxiliary load balancing.Open lesson
- 6State Space Models & MambaContinuous-time SSMs, selective scan parameters, discretization, and linear-time sequence modeling.Open lesson
- 7Model ServingHigh-throughput LLM hosting, KV caching, vLLM, speculative decoding, and quantization.Open lesson
Decisions this path should help you make
- Is the workload compute-, memory-, or communication-bound?
- Which quality loss is acceptable for lower latency or memory?
- How should prefill and decode capacity be planned separately?
- What release evidence and rollback target are required?