Dontopedia
source context

Part 365

ctx:discord/blah/watt-activation/part-365
kind: custommode: permissive83 statements42 distinct subjects

No external document is attached to this context. (Many contexts are pure organisational labels.)

Facts in this context

Grouped by subject. Each subject links to its full article.

Recv Bug Fix6 factsex:recv-bug-fix

changedU Right Equals Shift U Now
describedAscorrectness
improvesCorrectness
locatedInFast Path Recv Computation
matchesSemanticsOfPropagate Causal
presupposesExistenceOfFast Path Recv Computation

Strict Norm6 factsex:strict-norm

deemedUnnecessaryBelow
makesNoDifferenceParity
noReasonToPreferProduction
setTotrue
testedViaBenchmark
unnecessaryInProduction

Changes5 factsex:changes

includesNextConv1d Default Fast Path
includesNextTraining Scripts Updated
includesNextConv1d Depthwise Propagation
includesNextNormalization Finding
orderedSequentiallyRecv Bug Fix

Conv1d Depthwise Propagation5 factsex:conv1d-depthwise-propagation

deviatesFromShift Path
hasParityDeviation2e-7
maintainsParityShift Path
replacedConcat Shift Plus Kappa U Right Minus U Now Plus 2 Gamma U Now
usesMx Conv1d Groups C

Summary4 factsex:summary

listsActionsChanges
postedTwice2026-03-17 21:21
providesSummaryOfDone Actions
timestamped2026-03-17 21:21

Build Fast Forward3 factsex:build-fast-forward

hasParampropagation=conv1d
hasParamdtype=mx.bfloat16
hasParammodel

Config D24 D3843 factsex:config-d24-d384

baselineTokensPerSecond86K
fastConv1dBf16TokensPerSecond105K
speedup1.23×

Config D24 D7683 factsex:config-d24-d768

baselineTokensPerSecond77K
fastConv1dBf16TokensPerSecond88K
speedup1.14×

Config D48 D3843 factsex:config-d48-d384

baselineTokensPerSecond43K
fastConv1dBf16TokensPerSecond52K
speedup1.22×

Conv1d3 factsex:conv1d

assumesMxFramework
isDefaultFast Path
preferredOverShift

Kernel3 factsex:kernel

builtFromPhysics Params
hasValue[κ, 2-γ-κ]
reusedForK Steps

Benchmark2 factsex:benchmark

demonstratesSpeedup
showedStrict Norm Zero Difference to Parity

Benchmark Results2 factsex:benchmark-results

forRegimeThin Deep
presentedInTable

Propagation2 factsex:propagation

hasLegacyFallbackshift
hasOptionconv1d

Propagation Conv1d2 factsex:propagation-conv1d

enablesSpeedupBenchmark Results
resultsInFewer Metal Dispatches

Train Antenna Py2 factsex:train-antenna-py

acceptsFlagCompiled
updatedTraining Scripts

Xenonfun2 factsex:xenonfun

informsAudienceChanges
postedSummary

Backward Compat Aliases1 factex:backward-compat-aliases

preservedForImports

Bfloat161 factex:bfloat16

usedInFast Forward

Bug Fix1 factex:bug-fix

restoresOriginalSemantics

Compiled1 factex:--compiled

usesBuild Fast Forward

Curriculum Handling1 factex:curriculum-handling

preserved

Fast Conv1d Bf161 factex:fast-conv1d-bf16

speedupOverBaseline~1.2x

Fast Path1 factex:fast-path

optimizedForSpeed

Fp321 factex:fp32

usedInDeviation Measure

Kernel Reuse1 factex:kernel-reuse

optimizesPerformance

Metal Dispatches1 factex:metal-dispatches

reducedByConv1d

Mx Conv1d Groups C1 factex:mx-conv1d-groups-c

calledPerPropagation Step

Nan Guards1 factex:nan-guards

preserved

Original Forward1 factex:original-forward

providesUnbiased Metrics

Production1 factex:production

avoidsStrictNorm

Propagate Causal1 factex:_propagate_causal

computesSpatial Gradient

Spatial Gradient1 factex:spatial-gradient

computedOnField Input to Last Propagation

Thin Deep1 factex:thin-deep

describedAsbest regime

Thin Deep Regime1 factex:thin-deep-regime

evaluatedAsBest

Train Antenna Long Py1 factex:train-antenna-long-py

updatedTraining Scripts

Training Scripts1 factex:training-scripts

supportsCompiledMode

Unbiased Metrics1 factex:unbiased-metrics

achievedByOriginal Forward

U Right Equals Shift U Now1 factex:u-right-equals-shift-u-now

replacedByU Right Equals Shift U Prev

Validation1 factex:validation

usesOriginal Forward