| ⍝ LEGACY experiment, not the production runtime. Production arithmetic is j/biencoder.ijs and r/biencoder.R. |
| ⍝ TinyAPL Complex Double cannot hold Goldilocks p = 18446744069414584321. |
| ⍝ Mini SUBLEQ bi-encoder. NOT softmax attention. NOT a trained model. |
| ⍝ Query tower: Q ← A +/∙× Wq. Key tower: K ← (⍉B) +/∙× Wk. |
| ⍝ Late interaction calls library SubleqPipeline on each (Q row, K row): |
| ⍝ diff ← K−Q, pred ← diff≤0, then predicted dot ← +/ pred×Q×K. |
| ⍝ Unmasked head is Q +/∙× ⍉K. With the fixed permutation weights |
| ⍝ (Wq +/∙× ⍉Wk = I) that head equals A +/∙× B. The routed head is a |
| ⍝ sparse approximation; compare it to the reference matmul and record error. |
| ⍝ Full Goldilocks p is not an exact TinyAPL scalar. This file uses ordinary |
| ⍝ small integers plus GSafeP=65537 on the routed sum. Python is authoritative |
| ⍝ for p = 2^64−2^32+1. |
| LcgStep←{2147483647|48271×⍵} |
| LcgDraw←{n←⍺⋄seed←0⊇⍵⋄mod←1⊇⍵⋄buf←⟨⟩⋄i←0⋄{seed↩LcgStep seed⋄buf↩buf⍪mod|seed⋄i↩i+1⋄0}⍣{i≥n}0⋄⟨buf⋄seed⟩} |
| RoutedDot←{st←⍺ SubleqPipeline ⍵⋄pred←⊃2⊇st⋄+/(pred×⍺×⍵)} |
| RouteAgg←{0⊇⍺ SubleqPipeline ⍵} |
| RoutePred←{st←⍺ SubleqPipeline ⍵⋄⊃2⊇st} |
| PredictRows←{qq←↓⍺⋄kk←↓⍵⋄↑{q←⍵⋄{q RoutedDot ⍵}¨kk}¨qq} |
| AggRows←{qq←↓⍺⋄kk←↓⍵⋄↑{q←⍵⋄{q RouteAgg ⍵}¨kk}¨qq} |
| PredFlat←{qq←↓⍺⋄kk←↓⍵⋄∊{q←⍵⋄∊{q RoutePred ⍵}¨kk}¨qq} |
| GSafeRoutedDot←{st←⍺ SubleqPipeline ⍵⋄pred←⊃2⊇st⋄GSafeReduce pred×⍺×⍵} |
| GSafePredict←{qq←↓⍺⋄kk←↓⍵⋄↑{q←⍵⋄{q GSafeRoutedDot ⍵}¨kk}¨qq} |
|
|
