x54-729
commited on
Commit
•
90ff09a
1
Parent(s):
7f2ea77
remove unnecessary attention_drop
Browse files- modeling_internlm.py +1 -3
modeling_internlm.py
CHANGED
@@ -417,10 +417,8 @@ class InternLMFlashAttention2(InternLMAttention):
|
|
417 |
key_states = key_states.transpose(1, 2)
|
418 |
value_states = value_states.transpose(1, 2)
|
419 |
|
420 |
-
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
421 |
-
|
422 |
attn_output = self._flash_attention_forward(
|
423 |
-
query_states, key_states, value_states, attention_mask, q_len
|
424 |
)
|
425 |
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
426 |
attn_output = self.o_proj(attn_output)
|
|
|
417 |
key_states = key_states.transpose(1, 2)
|
418 |
value_states = value_states.transpose(1, 2)
|
419 |
|
|
|
|
|
420 |
attn_output = self._flash_attention_forward(
|
421 |
+
query_states, key_states, value_states, attention_mask, q_len
|
422 |
)
|
423 |
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
424 |
attn_output = self.o_proj(attn_output)
|