Logical parameters = actual model parameters (NOT packed tensor elements). For MXFP4, 2 FP4 values per U8 byte.

Active parameters/token = parameters used per forward pass. For MoE, only top_k experts + shared experts are active.

MoE (Mixture of Experts): routed experts are sharded by EP, shared experts are always active.

MLA (Multi-head Latent Attention): uses compressed latent KV (kv_lora_rank + qk_rope_head_dim), not full K/V.

MTP (Multi-Token Prediction): auxiliary module for speculative decoding. Excluded from normal inference unless enabled.

KV cache = per-token memory for cached keys/values. MLA uses latent formula; GQA/MQA use classic formula.

TP (Tensor Parallel): shards attention + non-expert weights across GPUs within one NVSwitch domain.

EP (Expert Parallel): shards routed expert weights across GPUs. TP=EP=8 means 8 GPUs, not 64.

PP (Pipeline Parallel): splits layers across stages. No latency penalty modeled yet.

KV pool = remaining VRAM after weights + overhead. Concurrency = KV pool / KV-per-token / P95-active-tokens.

Roofline = theoretical compute/memory lower bound. NOT realistic TTFT/TPOT — shown as UNKNOWN without benchmark.

Waste penalty: penalizes oversized GPUs. An 8B model on B300 (288GB) gets penalized vs H100 (80GB).

Safety margin: 3% of usable VRAM kept free. Margin > 0 for every PASS candidate.

References: DeepSeek-V3 GitHub, vLLM recipes, OpenAI gpt-oss docs, NVIDIA GPU specs. See docs/calculation-methodology.md for full details.