Why is inference slow?
Separate prefill, decode, KV cache reads, batching, and speculative decoding before blaming the model.
InfraLens
A clear starting point for learning AI infrastructure.
InfraLens helps you build a first map of memory, latency, parallelism, quantization, and token flow — with paths, formulas, calculators, and small code examples when you want to try things yourself.
Built for exploration, interview preparation, and technical research — not to cover everything, but to give you a clear place to start.
Entry Points
Start with the problem you are trying to explain.
Each card gives you a short route instead of dropping you into a long index.
Why does training not fit?
Break memory into weights, gradients, optimizer states, activations, and communication buffers.
Why does long context hurt?
Look at positional behavior, KV cache growth, attention cost, and serving latency as separate pressures.
Why is video generation expensive?
Follow the cost from image tokens to latent grids, temporal patches, denoising steps, and control paths.
Paths
Choose a learning path when you want structure.
Use these routes when you want a topic sequence rather than a concept lookup.
LLM InferenceLatency, KV cache, batching, and decode bottlenecks.
Distributed TrainingMemory states, sharding, and communication trade-offs.
Generative SystemsDiffusion, flow, latent tokens, and video cost.
Multimodal SystemsVision tokens, projection layers, and context pressure.
Agents and Post-TrainingPreference losses, reasoning, tools, and feedback loops.
Workspace
Use the right surface for the next question.
Concepts
Look up a mechanism, formula, pitfall, or code example in context.
Browse conceptsTools
Estimate KV cache, attention memory, video tokens, and quantized weights.
Try calculatorsMap
See how concepts connect when you know one term but not the surrounding system.
Open concept map