Upload 2 files
Browse files- configuration_minicpm.py +5 -0
- modeling_minicpm.py +4 -4
configuration_minicpm.py
CHANGED
|
@@ -174,6 +174,11 @@ class MiniCPMConfig(PretrainedConfig):
|
|
| 174 |
tie_word_embeddings=tie_word_embeddings,
|
| 175 |
**kwargs,
|
| 176 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 177 |
|
| 178 |
def _rope_scaling_validation(self):
|
| 179 |
"""
|
|
|
|
| 174 |
tie_word_embeddings=tie_word_embeddings,
|
| 175 |
**kwargs,
|
| 176 |
)
|
| 177 |
+
try:
|
| 178 |
+
import flash_attn
|
| 179 |
+
self._attn_implementation = "flash_attention_2"
|
| 180 |
+
except:
|
| 181 |
+
pass
|
| 182 |
|
| 183 |
def _rope_scaling_validation(self):
|
| 184 |
"""
|
modeling_minicpm.py
CHANGED
|
@@ -51,10 +51,11 @@ from transformers.utils.import_utils import is_torch_fx_available
|
|
| 51 |
from .configuration_minicpm import MiniCPMConfig
|
| 52 |
import re
|
| 53 |
|
| 54 |
-
|
| 55 |
-
if is_flash_attn_2_available():
|
| 56 |
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
| 57 |
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
|
|
|
|
|
|
| 58 |
|
| 59 |
|
| 60 |
# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
|
|
@@ -125,7 +126,7 @@ ALL_LAYERNORM_LAYERS.append(MiniCPMRMSNorm)
|
|
| 125 |
|
| 126 |
|
| 127 |
class MiniCPMRotaryEmbedding(nn.Module):
|
| 128 |
-
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=
|
| 129 |
super().__init__()
|
| 130 |
|
| 131 |
self.dim = dim
|
|
@@ -763,7 +764,6 @@ class MiniCPMDecoderLayer(nn.Module):
|
|
| 763 |
def __init__(self, config: MiniCPMConfig, layer_idx: int):
|
| 764 |
super().__init__()
|
| 765 |
self.hidden_size = config.hidden_size
|
| 766 |
-
|
| 767 |
self.self_attn = MINICPM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
|
| 768 |
|
| 769 |
self.mlp = MiniCPMMLP(config)
|
|
|
|
| 51 |
from .configuration_minicpm import MiniCPMConfig
|
| 52 |
import re
|
| 53 |
|
| 54 |
+
try:
|
|
|
|
| 55 |
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
| 56 |
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
| 57 |
+
except:
|
| 58 |
+
pass
|
| 59 |
|
| 60 |
|
| 61 |
# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
|
|
|
|
| 126 |
|
| 127 |
|
| 128 |
class MiniCPMRotaryEmbedding(nn.Module):
|
| 129 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
| 130 |
super().__init__()
|
| 131 |
|
| 132 |
self.dim = dim
|
|
|
|
| 764 |
def __init__(self, config: MiniCPMConfig, layer_idx: int):
|
| 765 |
super().__init__()
|
| 766 |
self.hidden_size = config.hidden_size
|
|
|
|
| 767 |
self.self_attn = MINICPM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
|
| 768 |
|
| 769 |
self.mlp = MiniCPMMLP(config)
|