Dontopedia

MLX

From Dontopedia, the open, paraconsistent wiki. (Last updated 2026-06-06.)

MLX has 98 facts recorded in Dontopedia across 52 references, with 5 live disagreements.

98 facts·80 predicates·52 sources·5 in dispute

Mostly:rdf:type(9), is primarily(3), handles cache cleanup(2)

Maturity scale raw canonical shape-checked rule-derived certified

Inbound mentions (47)

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.

usesFrameworkUses Framework(3)

targetFrameworkTarget Framework(2)

advocatesPortingToAdvocates Porting to(1)

affirmsComplexSupportAffirms Complex Support(1)

affirmsDenseLazyGraphDefaultAffirms Dense Lazy Graph Default(1)

affirmsLimitedSparseSupportAffirms Limited Sparse Support(1)

assumesKnowledgeOfAssumes Knowledge of(1)

avoidsAvoids(1)

bundlesNativelyBundles Natively(1)

canBeBuiltInCan Be Built in(1)

cannotRunOnCannot Run on(1)

claimsFeatureExistenceClaims Feature Existence(1)

demonstratesKnowledgeOfDemonstrates Knowledge of(1)

didFullMlxPipelineDid Full Mlx Pipeline(1)

excludesExcludes(1)

focusesOnFocuses on(1)

implementedInImplemented in(1)

involvesLibraryInvolves Library(1)

isCorrectForIs Correct for(1)

isOptimizedOnIs Optimized on(1)

likelyTransitiveViaLikely Transitive Via(1)

mentionsToolMentions Tool(1)

notPortedToNot Ported to(1)

notUsingNot Using(1)

notWorkingRightNot Working Right(1)

nowOptimizedNow Optimized(1)

optimizedForOptimized for(1)

optimizedForMlxOptimized for Mlx(1)

partOfPart of(1)

plansToAddChunkToPlans to Add Chunk to(1)

presupposesExistenceOfPresupposes Existence of(1)

proposesPortToProposes Port to(1)

referencesFrameworkReferences Framework(1)

referencesMlFrameworkReferences ML Framework(1)

referencesMlxFrameworkReferences Mlx Framework(1)

requiresFrameworkRequires Framework(1)

runsInRuns in(1)

standardPatternInStandard Pattern in(1)

statedIntentToPortStated Intent to Port(1)

suitableForPortingToSuitable for Porting to(1)

triggersOnTopicTriggers on Topic(1)

usesDirectScanLikeUses Direct Scan Like(1)

usesMlxFrameworkUses Mlx Framework(1)

wasNotOptimizedPreviouslyWas Not Optimized Previously(1)

Other facts (92)

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.

92 facts
PredicateValueRef
Rdf:typeSoftware Framework[41]
Rdf:typeSoftware Framework[42]
Rdf:typeFramework[44]
Rdf:typeSoftware Framework[45]
Rdf:typeSoftware Library[46]
Rdf:typeSoftware Library[47]
Rdf:typeSoftware Framework[49]
Rdf:typeSoftware Framework[50]
Rdf:typeFramework[51]
Is PrimarilyDense Tensor Compute[21]
Is PrimarilyDynamic Execution Optional Compilation[21]
Is PrimarilyLazy Graph Construction[21]
Handles Cache CleanupNegative Drift[8]
Handles Cache CleanupBetween Runs[8]
Is FrameworkApple ML[16]
Is Frameworknull[35]
Is Apple ML Frameworktrue[1]
Provides Fused Operationmx.fast.scaled_dot_product_attention[2]
Optimizes for Apple Hardwarenull[2]
Supports Transformer Trainingnull[2]
Used in ImplementationAdapter Trainer Py[3]
Lacks Native SupportPer Layer Groups[3]
Optimizes WellEinsum[4]
Context FrameworkCodebase[4]
Is LazyExecution[5]
Has Lazy Semanticsnull[5]
Requires Force forOptimizer State Updates[5]
References Apple Mlx FrameworkML Framework[6]
Has Mpi Supporteasy[7]
Supports Tb5 Link Aggregationtrue[7]
Provides Easy MpiStuff[7]
Supports Rdma Over Tb5true[7]
References Apple Mlx LibraryMlx Framework[6]
Has Mx CumsumMx.cumsum[9]
Lacks Py Torch Cuda Problemtrue[9]
Supports Parallel Cumsumtrue[9]
Apple ML FrameworkContext[10]
Is Target Porting Platformnull[11]
Provides Mx EvalMx Eval[12]
Provides Mx TopkMx Topk[12]
Handles Oob Silentlytrue[13]
Presupposes No Error on OobGarbage Return[13]
Contextual LibraryML Framework[13]
Does Handle Oob SilentlyReturns Garbage[13]
Returns Garbage Not ErrorSilent Oob[13]
Supports Value and Gradnull[14]
Hosts ImplementationCurrent Training Run[15]
Supports Side ChannelsSide Channel[17]
Contrasts WithCuda[18]
Framework Usednull[19]
Contextualizes Harmonic InferApple Mlx[20]
Presupposes Support for Complex Tensors Operatorsin general[21]
Lacks Primary Strength inSparse Graphs Tensors[21]
Built Mainly AroundDense Arrays Dense Kernels[21]
Supports Complex Numbersyes[21]
Has Optional Compilationtrue[21]
Teleologically Dense FocusedDense Kernels[21]
Has Lazy Computation Graphtrue[21]
Is ML Frameworknull[22]
Is Target FrameworkSpherical Vq[23]
Uses Lazy GraphsTransformer Forward[24]
Is Contextual FrameworkApple Mlx[25]
References Apple ML Frameworknull[26]
Preferred Over Python IntsMx Array[27]
Has Unusable Performancetrue[28]
Is Workaround Avoidedtrue[28]
References FrameworkApple Mlx[29]
Compatible With Numpynull[30]
Has PrimitivesCumsum[31]
Has Good Support forScan Compile[31]
Supports PrimitivesOfdm Wire Encoding S H 1[32]
Is Implementation Target{}[32]
Refers to FrameworkApple Mlx[33]
Performs Well on Dynamic Codenull[33]
Handles Compiled Many Timestrue[33]
Will Receive Chunk onGate Experiments[34]
Already Optimizes Fusionnull[35]
Has Lazy Evaluationtrue[35]
Used in BatchingSingle Numpy Mlx Call[36]
Used inHarmonic Mlx Parity Work[37]
Supports Cumsum ParallelMx Cumsum[38]
Contextually RelevantApple ML Framework[8]
Is Target PlatformPorting[39]
Is Suitable for ScalingSymbiogenesis[40]
Has Evaluation StrategyLazy Evaluation[43]
Lacks ProblemPyTorch/CUDA problem[44]
Has Functionmx.cumsum[44]
Has Cumsum OperationMx Cumsum[45]
Has Support forCuda[47]
Has FeatureFused Attention Kernel[48]
Has Capabilityhandles it many times compiled[51]
Runs on Gpu by Defaulttrue[52]

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.

98 undated
—
isAppleMLFrameworkblah/random/part-42
true
—
providesFusedOperationblah/watt-activation/part-20
mx.fast.scaled_dot_product_attention
—
optimizesForAppleHardwareblah/watt-activation/part-20
null
—
supportsTransformerTrainingblah/watt-activation/part-20
null
—
usedInImplementationblah/watt-activation/part-34
ex:adapter-trainer-py
—
lacksNativeSupportblah/watt-activation/part-34
ex:per-layer-groups
—
optimizesWellblah/watt-activation/part-78
ex:einsum
—
contextFrameworkblah/watt-activation/part-78
ex:codebase
—
isLazyblah/watt-activation/part-84
ex:execution
—
hasLazySemanticsblah/watt-activation/part-84
null
—
requiresForceForblah/watt-activation/part-84
ex:optimizer-state-updates
—
referencesAppleMlxFrameworkblah/watt-activation/part-77
ex:ml-framework
—
hasMpiSupportblah/watt-activation/part-94
easy
—
supportsTb5LinkAggregationblah/watt-activation/part-94
true
—
providesEasyMpiblah/watt-activation/part-94
ex:stuff
—
supportsRdmaOverTb5blah/watt-activation/part-94
true
—
handlesCacheCleanupblah/watt-activation/part-80
ex:negative-drift
—
referencesAppleMLXLibraryblah/watt-activation/part-77
ex:mlx-framework
—
hasMxCumsumblah/watt-activation/part-101
ex:mx.cumsum
—
lacksPyTorchCudaProblemblah/watt-activation/part-101
true
—
supportsParallelCumsumblah/watt-activation/part-101
true
—
appleMlFrameworkblah/watt-activation/part-104
ex:context
—
isTargetPortingPlatformblah/watt-activation/part-115
null
—
providesMxEvalblah/watt-activation/part-108
ex:mx-eval
—
providesMxTopkblah/watt-activation/part-108
ex:mx-topk
—
handlesOobSilentlyblah/watt-activation/part-119
true
—
presupposesNoErrorOnOobblah/watt-activation/part-119
ex:garbage-return
—
contextualLibraryblah/watt-activation/part-119
ex:ml-framework
—
doesHandleOobSilentlyblah/watt-activation/part-119
ex:returns-garbage
—
returnsGarbageNotErrorblah/watt-activation/part-119
ex:silent-oob
—
supportsValueAndGradblah/watt-activation/part-117
null
—
hostsImplementationblah/watt-activation/part-153
ex:current-training-run
—
isFrameworkblah/watt-activation/part-167
ex:apple-ml
—
supportsSideChannelsblah/watt-activation/part-180
ex:side-channel
—
contrastsWithblah/watt-activation/part-217
ex:cuda
—
frameworkUsedblah/watt-activation/part-244
null
—
contextualizesHarmonicInferblah/watt-activation/part-241
ex:apple-mlx
—
presupposesSupportForComplexTensorsOperatorsblah/watt-activation/part-263
in general
—
isPrimarilyblah/watt-activation/part-263
ex:dense-tensor-compute
—
isPrimarilyblah/watt-activation/part-263
ex:dynamic-execution-optional-compilation
—
isPrimarilyblah/watt-activation/part-263
ex:lazy-graph-construction
—
lacksPrimaryStrengthInblah/watt-activation/part-263
ex:sparse-graphs-tensors
—
builtMainlyAroundblah/watt-activation/part-263
ex:dense-arrays-dense-kernels
—
supportsComplexNumbersblah/watt-activation/part-263
yes
—
hasOptionalCompilationblah/watt-activation/part-263
true
—
teleologicallyDenseFocusedblah/watt-activation/part-263
ex:dense-kernels
—
hasLazyComputationGraphblah/watt-activation/part-263
true
—
isMlFrameworkblah/watt-activation/part-268
null
—
isTargetFrameworkblah/watt-activation/part-297
ex:spherical-vq
—
usesLazyGraphsblah/watt-activation/part-301
ex:transformer-forward
—
isContextualFrameworkblah/watt-activation/part-311
ex:apple-mlx
—
referencesAppleMLFrameworkblah/watt-activation/part-324
null
—
preferredOverPythonIntsblah/watt-activation/part-329
ex:mxArray
—
hasUnusablePerformanceblah/watt-activation/part-340
true
—
isWorkaroundAvoidedblah/watt-activation/part-340
true
—
referencesFrameworkblah/watt-activation/part-363
ex:apple-mlx
—
compatibleWithNumpyblah/watt-activation/part-364
null
—
hasPrimitivesblah/watt-activation/part-378
ex:cumsum
—
hasGoodSupportForblah/watt-activation/part-378
ex:scan-compile
—
supportsPrimitivesblah/watt-activation/part-384
ex:ofdm-wire-encoding-s-h-1
—
isImplementationTargetblah/watt-activation/part-384
{}
—
refersToFrameworkblah/watt-activation/part-385
ex:apple-mlx
—
performsWellOnDynamicCodeblah/watt-activation/part-385
null
—
handlesCompiledManyTimesblah/watt-activation/part-385
true
—
willReceiveChunkOnblah/watt-activation/part-412
ex:gate-experiments
—
alreadyOptimizesFusionblah/watt-activation/part-449
null
—
isFrameworkblah/watt-activation/part-449
null
—
hasLazyEvaluationblah/watt-activation/part-449
true
—
usedInBatchingblah/watt-activation/part-450
ex:single-numpy-mlx-call
—
usedInblah/watt-activation/part-477
ex:harmonic-mlx-parity-work
—
supportsCumsumParallelblah/watt-activation/part-103
ex:mx-cumsum
—
handlesCacheCleanupblah/watt-activation/part-80
ex:between-runs
—
contextuallyRelevantblah/watt-activation/part-80
ex:apple-ml-framework
—
isTargetPlatformblah/watt-activation/part-416
ex:porting
—
isSuitableForScalingblah/watt-activation/part-434
ex:symbiogenesis
—
typeblah/general/128
ex:SoftwareFramework
—
labelblah/general/128
MLX
—
labelblah/watt-activation/15
mlx
—
typeblah/watt-activation/15
ex:SoftwareFramework
—
labelblah/watt-activation/84
MLX
—
hasEvaluationStrategyblah/watt-activation/84
ex:lazy-evaluation
—
labelblah/watt-activation/101
MLX
—
typeblah/watt-activation/101
ex:Framework
—
lacksProblemblah/watt-activation/101
PyTorch/CUDA problem
—
hasFunctionblah/watt-activation/101
mx.cumsum
—
typeblah/watt-activation/103
ex:SoftwareFramework
—
hasCumsumOperationblah/watt-activation/103
ex:mx-cumsum
—
typeblah/watt-activation/119
ex:SoftwareLibrary
—
typeblah/watt-activation/266
ex:SoftwareLibrary
—
labelblah/watt-activation/266
mlx
—
hasSupportForblah/watt-activation/266
ex:cuda
—
hasFeatureblah/watt-activation/300
ex:fused-attention-kernel
—
typeblah/watt-activation/338
ex:SoftwareFramework
—
typeblah/watt-activation/382
ex:SoftwareFramework
—
typeblah/watt-activation/383
ex:Framework
—
labelblah/watt-activation/383
mlx
—
hasCapabilityblah/watt-activation/383
handles it many times compiled
—
runsOnGpuByDefaultblah/watt-activation/452
true

References (52)

52 references
  1. [1]Part 421 fact
    ctx:discord/blah/random/part-42
  2. [2]Part 203 facts
    ctx:discord/blah/watt-activation/part-20
  3. [3]Part 342 facts
    ctx:discord/blah/watt-activation/part-34
  4. [4]Part 782 facts
    ctx:discord/blah/watt-activation/part-78
  5. [5]Part 843 facts
    ctx:discord/blah/watt-activation/part-84
  6. [6]Part 772 facts
    ctx:discord/blah/watt-activation/part-77
  7. [7]Part 944 facts
    ctx:discord/blah/watt-activation/part-94
  8. [8]Part 803 facts
    ctx:discord/blah/watt-activation/part-80
  9. [9]Part 1013 facts
    ctx:discord/blah/watt-activation/part-101
  10. [10]Part 1041 fact
    ctx:discord/blah/watt-activation/part-104
  11. [11]Part 1151 fact
    ctx:discord/blah/watt-activation/part-115
  12. [12]Part 1082 facts
    ctx:discord/blah/watt-activation/part-108
  13. [13]Part 1195 facts
    ctx:discord/blah/watt-activation/part-119
  14. [14]Part 1171 fact
    ctx:discord/blah/watt-activation/part-117
  15. [15]Part 1531 fact
    ctx:discord/blah/watt-activation/part-153
  16. [16]Part 1671 fact
    ctx:discord/blah/watt-activation/part-167
  17. [17]Part 1801 fact
    ctx:discord/blah/watt-activation/part-180
  18. [18]Part 2171 fact
    ctx:discord/blah/watt-activation/part-217
  19. [19]Part 2441 fact
    ctx:discord/blah/watt-activation/part-244
  20. [20]Part 2411 fact
    ctx:discord/blah/watt-activation/part-241
  21. [21]Part 26310 facts
    ctx:discord/blah/watt-activation/part-263
  22. [22]Part 2681 fact
    ctx:discord/blah/watt-activation/part-268
  23. [23]Part 2971 fact
    ctx:discord/blah/watt-activation/part-297
  24. [24]Part 3011 fact
    ctx:discord/blah/watt-activation/part-301
  25. [25]Part 3111 fact
    ctx:discord/blah/watt-activation/part-311
  26. [26]Part 3241 fact
    ctx:discord/blah/watt-activation/part-324
  27. [27]Part 3291 fact
    ctx:discord/blah/watt-activation/part-329
  28. [28]Part 3402 facts
    ctx:discord/blah/watt-activation/part-340
  29. [29]Part 3631 fact
    ctx:discord/blah/watt-activation/part-363
  30. [30]Part 3641 fact
    ctx:discord/blah/watt-activation/part-364
  31. [31]Part 3782 facts
    ctx:discord/blah/watt-activation/part-378
  32. [32]Part 3842 facts
    ctx:discord/blah/watt-activation/part-384
  33. [33]Part 3853 facts
    ctx:discord/blah/watt-activation/part-385
  34. [34]Part 4121 fact
    ctx:discord/blah/watt-activation/part-412
  35. [35]Part 4493 facts
    ctx:discord/blah/watt-activation/part-449
  36. [36]Part 4501 fact
    ctx:discord/blah/watt-activation/part-450
  37. [37]Part 4771 fact
    ctx:discord/blah/watt-activation/part-477
  38. [38]Part 1031 fact
    ctx:discord/blah/watt-activation/part-103
  39. [39]Part 4161 fact
    ctx:discord/blah/watt-activation/part-416
  40. [40]Part 4341 fact
    ctx:discord/blah/watt-activation/part-434
  41. [41]1282 facts
    ctx:discord/blah/general/128
    • full textgeneral-128
      text/plain3 KBdoc:agent/general-128/c4588bce-f1f2-4a72-896f-b209a04b555e
      Show excerpt
      [2026-04-11 05:00] xenonfun: yeah Nemo Cascade 2 was quite good. at 63GB at full it was usable on the 96GB max tho I was screwing around and crashed machine, at 8-bit no issues. Gemma4 26B seemed to be sweet spot, tho was still a little ear…
  42. [42]152 facts
    ctx:discord/blah/watt-activation/15
    • full textwatt-activation-15
      text/plain3 KBdoc:agent/watt-activation-15/13ad2519-f6c2-47c1-afd8-14c1f26821f5
      Show excerpt
      [2026-03-03 03:48] xenonfun: some nocopy memory things still same speed, needs to start fusing kernels. [2026-03-03 04:30] xenonfun: Now I have a complete picture. Let me write the documentation first, then plan and implement the kernel fus…
  43. [43]842 facts
    ctx:discord/blah/watt-activation/84
    • full textwatt-activation-84
      text/plain3 KBdoc:agent/watt-activation-84/16e41088-c84d-4a6f-9c2d-56d69830cfa6
      Show excerpt
      [2026-03-07 20:41] xenonfun: okay some instant issues with this much data: ``` The problem: mx.eval(loss, model.parameters(), optimizer.state) traverses the full tree of 113M params + Adam's 2x state every step. For the compiled path, mx.ev…
  44. [44]1014 facts
    ctx:discord/blah/watt-activation/101
    • full textwatt-activation-101
      text/plain3 KBdoc:agent/watt-activation-101/6f90e9c2-2d77-484f-96fd-98a89c99440a
      Show excerpt
      [2026-03-08 18:08] xenonfun: ### Concrete comparison at T=4096, H=12, d_h=64: ResonanceV2 (cumsum, K=8 bands): Work: 12 × 8 × 4096 × 64 = 25M multiply-adds Kernel launches: ~K cumsums = 8 Metal kernels Memory: O(T · H · K · …
  45. [45]1032 facts
    ctx:discord/blah/watt-activation/103
    • full textwatt-activation-103
      text/plain3 KBdoc:agent/watt-activation-103/6d322edd-8b82-4859-be6f-bc7033a53fe1
      Show excerpt
      [2026-03-08 18:36] xenonfun: It appears your agents have actually already done all this work <@1211062099137265723> in your repo already. https://github.com/MonumentalSystems/harmonic-gpt/blob/master/docs/M3_DEPLOY_NOTES.md (files: Screens…
  46. [46]1191 fact
    ctx:discord/blah/watt-activation/119
    • full textwatt-activation-119
      text/plain3 KBdoc:agent/watt-activation-119/dd015076-4b38-4017-9483-3f91bdce858d
      Show excerpt
      [2026-03-09 00:25] xenonfun: okay at least generating something probablt still some bugs. ⏺ Committed and pushed. Key things done this session: 1. docs/symbiogenesis.md saved as a core document, linked prominently from CLAUDE.md 2. Roo…
  47. [47]2663 facts
    ctx:discord/blah/watt-activation/266
    • full textwatt-activation-266
      text/plain2 KBdoc:agent/watt-activation-266/0dd3318c-38a8-4ab0-8b7a-743748e72c54
      Show excerpt
      [2026-03-13 07:20] xenonfun: • Ran it. Long-prompt test (context_patches=128000, prompt = 1,024,000 bytes, generated 64 patches, compiled cached decode): - prompt bytes: 1,024,000 - generated patches: 64 - total elapsed: 231.8s …
  48. [48]3001 fact
    ctx:discord/blah/watt-activation/300
    • full textwatt-activation-300
      text/plain3 KBdoc:agent/watt-activation-300/3b6edccf-3524-4608-838f-25890efaea15
      Show excerpt
      [2026-03-14 06:34] xenonfun: ``` 3. Manual attention (lines 110-128) — Hand-rolled softmax attention instead of using mx.fast.scaled_dot_product_attention. MLX's fused attention kernel is significantly faster for small sequence lengths. …
  49. [49]3381 fact
    ctx:discord/blah/watt-activation/338
    • full textwatt-activation-338
      text/plain3 KBdoc:agent/watt-activation-338/5291b646-c08b-45ca-b1fe-b63fc86c3354
      Show excerpt
      [2026-03-15 16:56] xenonfun: ``` ⏺ No — LoheSphericalComplexAttention added complex gates (bandpass resonators) and complex coupling (phase-shifted sync). But the Lohe sync itself still normalizes to S^{H-1}: Q = lohe_normalize(self.proj…
  50. [50]3821 fact
    ctx:discord/blah/watt-activation/382
    • full textwatt-activation-382
      text/plain3 KBdoc:agent/watt-activation-382/7400f8cd-0f1a-470c-9260-c96918758eae
      Show excerpt
      [2026-03-19 01:37] xenonfun: ⏺ This is gold. Let me check the sweep results while I synthesize the architecture idea: ``` Here's the stripped-down principled design, pulling from the harmonic-gpt findings: "Resonant Wire" — Helmholtz+Cli…
  51. [51]3833 facts
    ctx:discord/blah/watt-activation/383
  52. [52]4521 fact
    ctx:discord/blah/watt-activation/452
    • full textwatt-activation-452
      text/plain3 KBdoc:agent/watt-activation-452/ff1dd4f5-3233-4ae2-8f83-249a90fd3e1d
      Show excerpt
      [2026-03-21 05:42] xenonfun: ⏺ The Rust timings include process startup + weight loading (~230ms), so let me subtract that overhead for a fair comparison: ``` ┌────────┬──────────────────┬──────────────────────────┬─────────┐ │ Tokens │…

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.

↑