Dontopedia

throughput gap

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-07-04.)

throughput gap is G=4, H=4 small model.

17 facts·15 predicates·6 sources·1 in dispute

Mostly:caused by(2), is0 22x(1), caused by python loops(1)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (1)

Other subjects in dontopedia point AT this entity as a value. These are inverse relationships — e.g. "X motherOf this subject" — and answer questions the forward facts can't. Grouped by predicate.

summarizesThroughputGapRootCauseSummarizes Throughput Gap Root Cause(1)

Other facts (16)

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.

16 facts
PredicateValueRef
Caused byKv Inp 4x Larger[2]
Caused byPython Head Loop[3]
Is0 22x0.22x[1]
Caused by Python Loopstrue[1]
Expected for Short Sequencestrue[1]
Is Expectedtrue[1]
Is Priority to Address Nextnull[2]
Attributed to Implementationnull[3]
Rdf:typePerformance Metric[4]
Has Value0.22[4]
Has ContextPyTorch benchmark[4]
Has Causeshort sequences with Python-level loops[4]
Has Percentage Value12[5]
Attributed toPython Head Loop[5]
DescriptionG=4, H=4 small model[6]
Future Behaviorat GPT-2 scale the gap will be smaller[6]

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.

is0-22xblah/watt-activation/part-48
0.22x
causedByPythonLoopsblah/watt-activation/part-48
true
expectedForShortSequencesblah/watt-activation/part-48
true
isExpectedblah/watt-activation/part-48
true
isPriorityToAddressNextblah/watt-activation/part-203
null
causedByblah/watt-activation/part-203
ex:kv-inp-4x-larger
causedByblah/watt-activation/part-451
ex:python-head-loop
attributedToImplementationblah/watt-activation/part-451
null
typeblah/watt-activation/48
ex:PerformanceMetric
labelblah/watt-activation/48
throughput gap
hasValueblah/watt-activation/48
0.22
hasContextblah/watt-activation/48
PyTorch benchmark
hasCauseblah/watt-activation/48
short sequences with Python-level loops
hasPercentageValueblah/watt-activation/449
12
attributedToblah/watt-activation/449
ex:python-head-loop
descriptiondocument/030a0648-513f-45a0-a9a4-618fb7d44634
G=4, H=4 small model
futureBehaviordocument/030a0648-513f-45a0-a9a4-618fb7d44634
at GPT-2 scale the gap will be smaller

References (6)

6 references
  1. [1]Part 484 facts
    ctx:discord/blah/watt-activation/part-48
  2. [2]Part 2032 facts
    ctx:discord/blah/watt-activation/part-203
  3. [3]Part 4512 facts
    ctx:discord/blah/watt-activation/part-451
  4. [4]485 facts
    ctx:discord/blah/watt-activation/48
    • full textwatt-activation-48
      text/plain3 KBdoc:agent/watt-activation-48/17fbe6b3-28dd-44cc-9e32-ece98c3e9a74
      Show excerpt
      [2026-03-07 06:49] xenonfun: ``` All 6 configs completed successfully — no crashes. Full summary: Complex Oscillator Sweep Results (20K iters, 8K vocab) ┌──────────────────┬───────────┬──────────┬────────────────────────────────┐ │
  5. [5]4492 facts
    ctx:discord/blah/watt-activation/449
    • full textwatt-activation-449
      text/plain2 KBdoc:agent/watt-activation-449/f6da4410-b0fa-46d2-9cb5-99d205dd2e55
      Show excerpt
      [2026-03-21 03:48] xenonfun: ⏺ You're right. At step 950: ``` ┌──────────┬───────────┬──────────┐ │ Metric │ Helmholtz │ v3 │ ├──────────┼───────────┼──────────┤ │ BPB │ 2.257 │ 2.257 │ ├──────────┼───────────┼
  6. ctx:claims/document/030a0648-513f-45a0-a9a4-618fb7d44634
    • full textxenonfun: Smoke test results (d=256, 4L, G=4, H=4): - r_global at random init: 0.69 (wit
      text/plain675 Bdiscord:convmsg/5853
      Show excerpt
      xenonfun: Smoke test results (d=256, 4L, G=4, H=4): - r_global at random init: 0.69 (within-block sync already present) - Mean r_global during training: 0.61 - Loss descent: 0.176 at 100 steps — matches spectral+lohe_v2 closely - Th

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.