README: KernelConfig usage with version/trust_remote_code
Browse files
README.md
CHANGED
|
@@ -45,7 +45,10 @@ import torch
|
|
| 45 |
from transformers import AutoModel, KernelConfig
|
| 46 |
|
| 47 |
kernel_config = KernelConfig({
|
| 48 |
-
"VJEPA2Encoder":
|
|
|
|
|
|
|
|
|
|
| 49 |
})
|
| 50 |
model = AutoModel.from_pretrained(
|
| 51 |
"facebook/vjepa2-vitl-fpc64-256",
|
|
@@ -107,7 +110,7 @@ exhaustion; details in the source repository's cycle records.
|
|
| 107 |
* **Hardware**: trn2 (verified on trn2.3xlarge, LNC=2).
|
| 108 |
* **SDK**: neuronx-cc 2.27.2878.0; torch-neuronx 2.11.3 (PyTorch Native); nki 0.6.0;
|
| 109 |
nkilib (bundled with the SDK — `ModularAllocator` is imported from it).
|
| 110 |
-
* **Transformers**
|
| 111 |
* `NKI_ENABLE_TRACE_CACHE=0` in the environment (the runtime defaults the cross-process trace
|
| 112 |
cache on; a stale NEFF silently captures a different instruction stream).
|
| 113 |
|
|
|
|
| 45 |
from transformers import AutoModel, KernelConfig
|
| 46 |
|
| 47 |
kernel_config = KernelConfig({
|
| 48 |
+
"VJEPA2Encoder": (
|
| 49 |
+
"aws-neuron/vjepa2-neuron-kernels:NeuronVJEPA2Encoder",
|
| 50 |
+
{"version": 1, "trust_remote_code": True},
|
| 51 |
+
),
|
| 52 |
})
|
| 53 |
model = AutoModel.from_pretrained(
|
| 54 |
"facebook/vjepa2-vitl-fpc64-256",
|
|
|
|
| 110 |
* **Hardware**: trn2 (verified on trn2.3xlarge, LNC=2).
|
| 111 |
* **SDK**: neuronx-cc 2.27.2878.0; torch-neuronx 2.11.3 (PyTorch Native); nki 0.6.0;
|
| 112 |
nkilib (bundled with the SDK — `ModularAllocator` is imported from it).
|
| 113 |
+
* **Transformers** `>= 5.14` (tuple-form `KernelConfig`); `kernels >= 0.17`.
|
| 114 |
* `NKI_ENABLE_TRACE_CACHE=0` in the environment (the runtime defaults the cross-process trace
|
| 115 |
cache on; a stale NEFF silently captures a different instruction stream).
|
| 116 |
|