v1.1: README pointer to vit-encoder-neuron-kernels (docs only, no code change)
Browse files- CHANGELOG.md +7 -0
- README.md +5 -0
CHANGELOG.md
CHANGED
|
@@ -1,5 +1,12 @@
|
|
| 1 |
# Changelog
|
| 2 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
## v1 (2026-09-25)
|
| 4 |
|
| 5 |
First release. Four NKI kernels replacing the per-layer torch ops of the V-JEPA2 ViT-L video
|
|
|
|
| 1 |
# Changelog
|
| 2 |
|
| 3 |
+
## v1.1 (2026-10)
|
| 4 |
+
|
| 5 |
+
Documentation only: README points to
|
| 6 |
+
[`jburtoft/vit-encoder-neuron-kernels`](https://huggingface.co/kernels/jburtoft/vit-encoder-neuron-kernels),
|
| 7 |
+
the generalized successor (batching, LNC=1, fused epilogues, SwiGLU, DINOv3). No code change; the v1
|
| 8 |
+
kernels and layer are byte-identical.
|
| 9 |
+
|
| 10 |
## v1 (2026-09-25)
|
| 11 |
|
| 12 |
First release. Four NKI kernels replacing the per-layer torch ops of the V-JEPA2 ViT-L video
|
README.md
CHANGED
|
@@ -23,6 +23,11 @@ tags:
|
|
| 23 |
implementation this was verified against, and rel_l2 4.2e-2 vs the fp32 torch encoder (the
|
| 24 |
model's own bf16-matmul sensitivity).
|
| 25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
## What ships
|
| 27 |
|
| 28 |
| Component | Purpose |
|
|
|
|
| 23 |
implementation this was verified against, and rel_l2 4.2e-2 vs the fp32 torch encoder (the
|
| 24 |
model's own bf16-matmul sensitivity).
|
| 25 |
|
| 26 |
+
> **See also: [`jburtoft/vit-encoder-neuron-kernels`](https://huggingface.co/kernels/jburtoft/vit-encoder-neuron-kernels)**,
|
| 27 |
+
> a generalized version of these kernels. It adds batching (any batch size), LNC=1 (8-core) support,
|
| 28 |
+
> fused residual + LayerNorm epilogues, a SwiGLU FFN kernel, both RoPE pairings, and a DINOv3 layer
|
| 29 |
+
> alongside a V-JEPA2 one. This repo is unchanged and keeps its verified 16-frame / batch-1 path.
|
| 30 |
+
|
| 31 |
## What ships
|
| 32 |
|
| 33 |
| Component | Purpose |
|