xai-studies / studies /shared /GLOSSARY.md
lyffseba's picture
xai-studies 20260723T201820Z
2353a55 verified
|
Raw
History Blame Contribute Delete
2.73 kB

Glossary

Shared vocabulary across packages. Prefer package derived/SPEC.md for numbers.

Term Meaning
MoE Mixture of Experts — sparse FFN; only some experts run per token
Active params Parameters used for a typical forward (not total stored weights)
Shared expert Expert(s) always on, often combined with routed experts
Shared-expert sink Shared experts share normalized routing mass with top-k (Inkling)
Shared-expert add Shared expert output added after routed experts (Laguna) — not sink mass
Sigmoid router Token-choice scores via σ(logits), not softmax (Laguna LagunaTopKRouter)
Softplus attention gate Per-head (or per-element) softplus gate on attn output before o_proj (Laguna)
QK-norm RMSNorm on Q and K per head before RoPE (Laguna)
YaRN Yet another RoPE extensioN — long-context RoPE scaling (Laguna global layers)
OpenMDW Open Model Deep Wide license family (Laguna S 2.1: OpenMDW-1.1)
pool Poolside agent harness; native training/eval environment for Laguna
DFlash Poolside speculative draft model pair for Laguna serving
poolside_v1 Tool-call + reasoning parser name for vLLM/SGLang Laguna serve
Quantile Balancing Expert load balance from router-score quantiles (Kimi K3 claim)
Stable LatentMoE K3 named MoE framework at 16/896 sparsity
SWA Sliding-window attention — local context only
MLA Multi-Head Latent Attention — compressed KV (DeepSeek-style; K2.x)
KDA Kimi Delta Attention — linear/hybrid attention for long sequences
AttnRes Attention Residuals — selective depth retrieval (K3)
Relative PE Position via attention bias vs distance, not RoPE (Inkling)
RoPE Rotary positional embeddings (common baseline)
SConv Short causal convolution on residual branches (Inkling)
hMLP Hierarchical MLP patch encoder for images (Inkling)
dMel Discrete mel-bin audio tokens (Inkling)
MTP Multi-token prediction / speculative draft heads
μP / width multiplier Logit scaling by model width factor (Inkling logits_mup_width_multiplier)
QAT Quantization-aware training
Thinking effort Controllable test-time compute / CoT budget
Preserved thinking Multi-turn must return prior reasoning content (K3 / Laguna preserve_thinking)
Harness overfitting Model recalls native tool schemas instead of third-party definitions (Laguna limit)
Tinker Thinking Machines fine-tune platform
Primary source Official blog, model card, Hub config, paper — not secondary recaps