[Llama + Mistral] Add attention dropout (#27315)

* add droppouts

* add the dropout

* add doc in the config

* nits

* fix mistral config

* nits

src/transformers/models/llama/configuration_llama.py

@@ -95,7 +95,8 @@ class LlamaConfig(PretrainedConfig):
             experimental feature, subject to breaking API changes in future versions.
         attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
             Whether to use a bias in the query, key, value and output projection layers during self-attention.
-
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
 
     ```python
     >>> from transformers import LlamaModel, LlamaConfig
@@ -133,6 +134,7 @@ class LlamaConfig(PretrainedConfig):
         rope_theta=10000.0,
         rope_scaling=None,
         attention_bias=False,
+        attention_dropout=0.0,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -156,6 +158,7 @@ class LlamaConfig(PretrainedConfig):
         self.rope_scaling = rope_scaling
         self._rope_scaling_validation()
         self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
 
         super().__init__(
             pad_token_id=pad_token_id,

src/transformers/models/llama/modeling_llama.py

@@ -278,6 +278,7 @@ class LlamaAttention(nn.Module):
     def __init__(self, config: LlamaConfig):
         super().__init__()
         self.config = config
+        self.attention_dropout = config.attention_dropout
         self.hidden_size = config.hidden_size
         self.num_heads = config.num_attention_heads
         self.head_dim = self.hidden_size // self.num_heads
@@ -292,6 +293,7 @@ class LlamaAttention(nn.Module):
                 f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                 f" and `num_heads`: {self.num_heads})."
             )
+
         self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
         self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
         self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
@@ -404,6 +406,7 @@ class LlamaAttention(nn.Module):
 
         # upcast attention to fp32
         attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
         attn_output = torch.matmul(attn_weights, value_states)
 
         if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
@@ -489,10 +492,7 @@ class LlamaFlashAttention2(LlamaAttention):
         key_states = key_states.transpose(1, 2)
         value_states = value_states.transpose(1, 2)
 
-        # TODO: llama does not have dropout in the config??
-        # It is recommended to use dropout with FA according to the docs
-        # when training.
-        dropout_rate = 0.0  # if not self.training else self.attn_dropout
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
 
         # In PEFT, usually we cast the layer norms in float32 for training stability reasons
         # therefore the input hidden states gets silently casted in float32. Hence, we need

src/transformers/models/mistral/configuration_mistral.py

@@ -82,7 +82,8 @@ class MistralConfig(PretrainedConfig):
             The base period of the RoPE embeddings.
         sliding_window (`int`, *optional*, defaults to 4096):
             Sliding window attention window size. If not specified, will default to `4096`.
-
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
 
     ```python
     >>> from transformers import MistralModel, MistralConfig
@@ -119,6 +120,7 @@ class MistralConfig(PretrainedConfig):
         tie_word_embeddings=False,
         rope_theta=10000.0,
         sliding_window=4096,
+        attention_dropout=0.0,
         **kwargs,
     ):
         self.vocab_size = vocab_size
@@ -139,6 +141,7 @@ class MistralConfig(PretrainedConfig):
         self.rms_norm_eps = rms_norm_eps
         self.use_cache = use_cache
         self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
 
         super().__init__(
             pad_token_id=pad_token_id,

src/transformers/models/mistral/modeling_mistral.py

@@ -205,6 +205,7 @@ class MistralAttention(nn.Module):
         self.max_position_embeddings = config.max_position_embeddings
         self.rope_theta = config.rope_theta
         self.is_causal = True
+        self.attention_dropout = config.attention_dropout
 
         if (self.head_dim * self.num_heads) != self.hidden_size:
             raise ValueError(
@@ -284,6 +285,7 @@ class MistralAttention(nn.Module):
 
         # upcast attention to fp32
         attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
         attn_output = torch.matmul(attn_weights, value_states)
 
         if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
@@ -390,11 +392,7 @@ class MistralFlashAttention2(MistralAttention):
         # repeat k/v heads if n_kv_heads < n_heads
         key_states = repeat_kv(key_states, self.num_key_value_groups)
         value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        # TODO: Mistral does not have dropout in the config??
-        # It is recommended to use dropout with FA according to the docs
-        # when training.
-        dropout_rate = 0.0  # if not self.training else self.attn_dropout
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
 
         # In PEFT, usually we cast the layer norms in float32 for training stability reasons
         # therefore the input hidden states gets silently casted in float32. Hence, we need