Tracks

How GPUs execute parallel workloads — from threads and warps to tensor cores and FlashAttention.

• Why GPUs & CUDA software stack
• Execution model (threads, warps, SMs)
• Memory hierarchy (registers → HBM)
• Roofline model
• Memory access patterns & coalescing
• Tiling & matrix multiply
• Tensor cores & mixed precision
• Operator fusion & FlashAttention
• Triton & torch.compile

From tokenization to PagedAttention — how large language models process text and generate output.

• Tokenization (BPE)
• Embeddings & positional encoding
• Self-attention (Q, K, V)
• Transformer block
• Text generation & sampling
• KV cache
• Quantization (GPTQ, AWQ, GGUF)
• Batching (static vs continuous)
• PagedAttention

How vLLM, SGLang, and TensorRT-LLM actually serve LLMs — scheduler, memory, and serving-engine internals.

• Inference engine internals (vLLM)
• Speculative decoding
• Prefill/decode disaggregation
• Serving metrics & SLOs (TTFT, TPOT, P99)
• CUDA Graphs
• Multi-LoRA serving
• Prefix caching & RadixAttention

Foundations of AI agents — the loop, tools, workflows, retrieval, context engineering, planning, evals, and security — through visual interactive simulations.

• The agent loop & state
• Tool use & function calling (incl. MCP, Skills)
• Workflow patterns + subagent topology
• Retrieval & RAG
• Context engineering
• Planning & reflection
• Evals & diagnostics
• Security & the lethal trifecta
• Capstone: three designs, same task

Production agent engineering — durable harnesses, observability, layered guardrails, deployment, incident response, and running an agent fleet against an SLO.

• Production harness architecture
• Observability for agents
• Layered guardrails
• Cost & latency engineering
• Production evals & shadow mode
• Deployment & rollout
• Incident handling
• Agent teams (multi-agent orchestration)
• Capstone: reliability operations