Spaces:
Running
on
A100
Running
on
A100
MekkCyber
commited on
Commit
·
23210c5
1
Parent(s):
29baea6
updating quantization
Browse files
app.py
CHANGED
|
@@ -74,9 +74,9 @@ def quantize_model(model_name, quantization_type, group_size=128, auth_token=Non
|
|
| 74 |
else :
|
| 75 |
quantization_config = TorchAoConfig(quantization_type)
|
| 76 |
if device == "cuda" :
|
| 77 |
-
model = load_model_gpu(model_name, quantization_config, auth_token)
|
| 78 |
else :
|
| 79 |
-
model = load_model_cpu(model_name, quantization_config, auth_token)
|
| 80 |
|
| 81 |
return model
|
| 82 |
|
|
|
|
| 74 |
else :
|
| 75 |
quantization_config = TorchAoConfig(quantization_type)
|
| 76 |
if device == "cuda" :
|
| 77 |
+
model = load_model_gpu(model_name, quantization_config=quantization_config, auth_token=auth_token)
|
| 78 |
else :
|
| 79 |
+
model = load_model_cpu(model_name, quantization_config=quantization_config, auth_token=auth_token)
|
| 80 |
|
| 81 |
return model
|
| 82 |
|