x54-729
commited on
Commit
•
75073b2
1
Parent(s):
8542548
remove unnecessary attention_drop
Browse files- modeling_internlm2.py +1 -3
modeling_internlm2.py
CHANGED
@@ -480,10 +480,8 @@ class InternLM2FlashAttention2(InternLM2Attention):
|
|
480 |
key_states = key_states.transpose(1, 2)
|
481 |
value_states = value_states.transpose(1, 2)
|
482 |
|
483 |
-
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
484 |
-
|
485 |
attn_output = self._flash_attention_forward(
|
486 |
-
query_states, key_states, value_states, attention_mask, q_len
|
487 |
)
|
488 |
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
489 |
attn_output = self.wo(attn_output)
|
|
|
480 |
key_states = key_states.transpose(1, 2)
|
481 |
value_states = value_states.transpose(1, 2)
|
482 |
|
|
|
|
|
483 |
attn_output = self._flash_attention_forward(
|
484 |
+
query_states, key_states, value_states, attention_mask, q_len
|
485 |
)
|
486 |
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
487 |
attn_output = self.wo(attn_output)
|