Update README.md
Browse filesA series of ablations asking what’s actually necessary in transformer attention. Each step removed something thought to be essential. Nothing broke. It got better.
README.md
CHANGED
|
@@ -1,3 +1,17 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
---
|
| 4 |
+
Key Result
|
| 5 |
+
|
| 6 |
+
O(N) learned causal convolution beats O(N²) softmax attention on both perplexity AND throughput, with the advantage growing at longer sequences:
|
| 7 |
+
|
| 8 |
+
Model PPL Change TPS (128) TPS (2048) Speedup
|
| 9 |
+
Learned Conv O(N) 8.08 -3.2% 378,066 1,009,622 5.5x
|
| 10 |
+
Standard QKV O(N²) 8.34 baseline 317,968 183,408 1.0x
|
| 11 |
+
At 2048 tokens, the O(N) model is 5.5x faster while achieving better perplexity. The gap widens with sequence length because O(N) scales linearly while O(N²) scales quadratically.
|
| 12 |
+
|
| 13 |
+
https://github.com/MikeyBeez/DifferentialLR
|
| 14 |
+
https://medium.com/p/6659a3793322
|
| 15 |
+
https://doi.org/10.5281/zenodo.18498944
|
| 16 |
+
|
| 17 |
+
|