Update modeling_mpt.py
Browse filesThis PR adds the accelerate support for MPT models so that any user could load these models in 8bit / 4bit
- modeling_mpt.py +1 -0
modeling_mpt.py
CHANGED
|
@@ -23,6 +23,7 @@ Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
|
|
| 23 |
class MPTPreTrainedModel(PreTrainedModel):
|
| 24 |
config_class = MPTConfig
|
| 25 |
base_model_prefix = 'model'
|
|
|
|
| 26 |
|
| 27 |
class MPTModel(MPTPreTrainedModel):
|
| 28 |
|
|
|
|
| 23 |
class MPTPreTrainedModel(PreTrainedModel):
|
| 24 |
config_class = MPTConfig
|
| 25 |
base_model_prefix = 'model'
|
| 26 |
+
_no_split_modules=["MPTBlock"]
|
| 27 |
|
| 28 |
class MPTModel(MPTPreTrainedModel):
|
| 29 |
|