README: point to aws-neuron/vjepa2-neuron-kernels (repo moved to aws-neuron org; docs only)
Browse files
README.md
CHANGED
|
@@ -93,7 +93,7 @@ import types, torch
|
|
| 93 |
from kernels import get_kernel
|
| 94 |
from transformers import AutoModel
|
| 95 |
|
| 96 |
-
k = get_kernel("
|
| 97 |
|
| 98 |
model = AutoModel.from_pretrained("facebook/dinov3-vitl16-pretrain-lvd1689m",
|
| 99 |
dtype=torch.float32).eval().to("neuron")
|
|
@@ -205,7 +205,7 @@ build/torch-neuron/
|
|
| 205 |
|
| 206 |
Apache-2.0 (see `LICENSE`, `NOTICE`). The attention, MLP and residual/LayerNorm kernels derive from
|
| 207 |
the V-JEPA2 encoder kernels in
|
| 208 |
-
[`
|
| 209 |
developed by Antonio Mena ([@menaman123](https://huggingface.co/menaman123)). This repo generalizes
|
| 210 |
them: both RoPE pairings, any head count, head_dim 128, batching, padded-key skipping, an LNC=1
|
| 211 |
all-group variant, and the fused residual/LayerNorm epilogues. `swiglu_stream.py` is new.
|
|
|
|
| 93 |
from kernels import get_kernel
|
| 94 |
from transformers import AutoModel
|
| 95 |
|
| 96 |
+
k = get_kernel("aws-neuron/vit-encoder-neuron-kernels", version=1, trust_remote_code=True)
|
| 97 |
|
| 98 |
model = AutoModel.from_pretrained("facebook/dinov3-vitl16-pretrain-lvd1689m",
|
| 99 |
dtype=torch.float32).eval().to("neuron")
|
|
|
|
| 205 |
|
| 206 |
Apache-2.0 (see `LICENSE`, `NOTICE`). The attention, MLP and residual/LayerNorm kernels derive from
|
| 207 |
the V-JEPA2 encoder kernels in
|
| 208 |
+
[`aws-neuron/vjepa2-neuron-kernels`](https://huggingface.co/kernels/aws-neuron/vjepa2-neuron-kernels),
|
| 209 |
developed by Antonio Mena ([@menaman123](https://huggingface.co/menaman123)). This repo generalizes
|
| 210 |
them: both RoPE pairings, any head count, head_dim 128, batching, padded-key skipping, an LNC=1
|
| 211 |
all-group variant, and the fused residual/LayerNorm epilogues. `swiglu_stream.py` is new.
|