Animated walkthrough of Kimi Delta Attention (KDA) in Kimi-K3. Configure num_heads, head_dim, hidden_size, and TP to recompute every tensor shape and dtype live, then step through the four execution paths — prefill (chunk/FlashKDA), decode (fused_recurrent), fused decode (single CUDA kernel), and spec decode. Highlights the conv_state and recurrent_state caches as data flows through fused input projection → conv1d → delta-rule → output gate.

Closed-form diagnostic for LLM KV-cache tiers (L1 device / L2 host / L3 store). From hardware, model, and multi-turn workload, decide whether serving is compute-bound or KV-movement-bound via two hit rates — theoretical (1 − new/total_len) vs fill-threshold (1 − chunk/P). Computes the L1 KV pool, useful vs useless L1, prefill/decode max-req, load-back overlap ratio, and where L3 pays off. Distilled from the L3 KV-tier study.

Interactive simulation of Prefill-Decode disaggregated LLM serving with DEP (Data-Expert Parallelism). Visualize producer-consumer queue dynamics, configure prefill/decode node counts, DEP ranks, TPOT, and request parameters. Real-time charts show queue depths and decode utilization over time.

Discrete-event simulation of multi-turn agent serving with PD disaggregation, UMBP KV prefetch vs recompute decisions, tiered memory offload (HBM/DRAM/SSD), and animated particle visualization of request and KV data flows. Includes embedded break-even cost model formulas.

Interactive visualization of throughput–latency Pareto trade-offs for multi-turn agent workloads. Explore how concurrent sessions, inter-turn pause time, prefix caching, and context growth reshape the serving performance frontier. Based on closed-loop queuing theory and operator-level performance modeling inspired by AIConfigurator.

Interactive IEEE 754 floating-point visualizer. Click individual bits to see how sign, exponent, and mantissa fields combine to represent FP64, FP32, FP16, BF16, FP8, and FP4 (MXFP4/NVFP4 with block scale) numbers. Useful for understanding quantization trade-offs in LLM inference.