Step-mode inference
From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)
Step-mode inference has 17 facts recorded in Dontopedia across 3 references, with 4 live disagreements.
Mostly:rdf:type(2), pertains to component(2), involves method(2)
Maturity scale
raw canonical shape-checked rule-derived certifiedOther facts (15)
The long tail: predicates that appear too rarely to warrant their own section. Filter or scroll to find a specific one. Each row links to its source.
| Predicate | Value | Ref |
|---|---|---|
| Rdf:type | Code Change | [1] |
| Rdf:type | Code Change | [2] |
| Pertains to Component | Kan | [1] |
| Pertains to Component | Sync Attention | [1] |
| Involves Method | UnifiedRotorBlock::step | [3] |
| Involves Method | UnifiedRotorLM::step | [3] |
| Describes Transformation | Normalization Move Transformation | [1] |
| Affects File | Rotor Attention V2 File | [2] |
| Description Is | Removed 3 redundant cv reconstructions and a full v_tan² rescan | [2] |
| Involves Class | UnifiedRotorInferenceState | [3] |
| Has Initial Speed | 17 | [3] |
| Has Final Speed | 860 | [3] |
| Has Speed Unit | tok/s | [3] |
| Has Speedup Factor | 50 | [3] |
| Has Approximate Speedup | true | [3] |
Timeline
Timeline axis is valid_time — when each source says the fact was true in the world, not when Dontopedia learned about it. Retracted rows are kept for provenance; coloured stripes indicate the context kind.
References (3)
ctx:discord/blah/watt-activation/118- full textwatt-activation-118text/plain3 KB
doc:agent/watt-activation-118/ed79098d-1144-44f5-9941-e6b2b9c1caa7Show excerpt
[2026-03-08 23:43] xenonfun: Code Changes (3 important patterns) 1. Fused QKV projection in SpectralAttention - Separate q_proj, k_proj, v_proj → single qkv_proj = Linear(d_model, 3 * d_model). One matmul instead of three. We should po…
ctx:discord/blah/watt-activation/664ctx:discord/blah/watt-activation/681- full textwatt-activation-681text/plain3 KB
doc:agent/watt-activation-681/376be3cc-9082-437f-bd74-a16443059d2aShow excerpt
[2026-04-22 15:24] xenonfun: ``` Code changes this session: 1. GPU training path wired (train_step_gpu_cached via ResidentBackwardCache) — 2.49× vs CPU 2. Step-mode inference (UnifiedRotorBlock::step + UnifiedRotorLM::step + UnifiedRoto…
See also
Keep researching
Missing something or suspicious of what's here? Kick off a research session — a Claude agent will investigate, cite its sources, and file new facts into a dedicated context you can review before accepting into the shared view.