Upload endpoint.json with huggingface_hub
Browse files- endpoint.json +26 -0
endpoint.json
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"compute": {
|
| 3 |
+
"accelerator": "gpu",
|
| 4 |
+
"instanceSize": "medium",
|
| 5 |
+
"instanceType": "nvidia-a10g",
|
| 6 |
+
"scaling": {
|
| 7 |
+
"minReplica": 0,
|
| 8 |
+
"maxReplica": 1
|
| 9 |
+
}
|
| 10 |
+
},
|
| 11 |
+
"model": {
|
| 12 |
+
"framework": "pytorch",
|
| 13 |
+
"task": "text-generation",
|
| 14 |
+
"image": {
|
| 15 |
+
"huggingface": {}
|
| 16 |
+
}
|
| 17 |
+
},
|
| 18 |
+
"env": {
|
| 19 |
+
"PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
|
| 20 |
+
"CUDA_LAUNCH_BLOCKING": "1",
|
| 21 |
+
"TORCH_USE_CUDA_DSA": "1",
|
| 22 |
+
"TRANSFORMERS_OFFLINE": "0",
|
| 23 |
+
"HF_HUB_ENABLE_HF_TRANSFER": "1",
|
| 24 |
+
"MODEL_LOAD_TIMEOUT": "600"
|
| 25 |
+
}
|
| 26 |
+
}
|