StableLM: Fix dropout argument type error (#29236)

* fix stablelm dropout argument type error

* fix docs of _flash_attention_forward

* fix all docs of _flash_attention_forward

* fix docs of _flash_attention_forward in starcoder2

---------
Co-authored-by: oliang <oliang@tencent.com>

src/transformers/models/bark/modeling_bark.py

@@ -306,7 +306,7 @@ class BarkSelfFlashAttention2(BarkSelfAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/bart/modeling_bart.py

@@ -430,7 +430,7 @@ class BartFlashAttention2(BartAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/distilbert/modeling_distilbert.py

@@ -370,7 +370,7 @@ class DistilBertFlashAttention2(MultiHeadSelfAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/falcon/modeling_falcon.py

@@ -657,7 +657,7 @@ class FalconFlashAttention2(FalconAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/gemma/modeling_gemma.py

@@ -410,7 +410,7 @@ class GemmaFlashAttention2(GemmaAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py

@@ -425,7 +425,7 @@ class GPTBigCodeFlashAttention2(GPTBigCodeAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -407,7 +407,7 @@ class GPTNeoFlashAttention2(GPTNeoSelfAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -439,7 +439,7 @@ class GPTNeoXFlashAttention2(GPTNeoXAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/llama/modeling_llama.py

@@ -518,7 +518,7 @@ class LlamaFlashAttention2(LlamaAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/mbart/modeling_mbart.py

@@ -420,7 +420,7 @@ class MBartFlashAttention2(MBartAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/mistral/modeling_mistral.py

@@ -496,7 +496,7 @@ class MistralFlashAttention2(MistralAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/mixtral/modeling_mixtral.py

@@ -574,7 +574,7 @@ class MixtralFlashAttention2(MixtralAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/opt/modeling_opt.py

@@ -394,7 +394,7 @@ class OptFlashAttention2(OPTAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/phi/modeling_phi.py

@@ -540,7 +540,7 @@ class PhiFlashAttention2(PhiAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/qwen2/modeling_qwen2.py

@@ -502,7 +502,7 @@ class Qwen2FlashAttention2(Qwen2Attention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/stablelm/modeling_stablelm.py

@@ -549,7 +549,7 @@ class StableLmFlashAttention2(StableLmAttention):
         key_states = key_states.transpose(1, 2)
         value_states = value_states.transpose(1, 2)
 
-        dropout_rate = self.attention_dropout if self.training else 0.0
+        dropout_rate = self.attention_dropout.p if self.training else 0.0
 
         attn_output = self._flash_attention_forward(
             query_states,
@@ -586,7 +586,7 @@ class StableLmFlashAttention2(StableLmAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/starcoder2/modeling_starcoder2.py

@@ -481,7 +481,7 @@ class Starcoder2FlashAttention2(Starcoder2Attention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)

src/transformers/models/whisper/modeling_whisper.py

@@ -536,7 +536,7 @@ class WhisperFlashAttention2(WhisperAttention):
             attention_mask (`torch.Tensor`):
                 The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                 position of padding tokens and 1 for the position of non-padding tokens.
-            dropout (`int`, *optional*):
+            dropout (`float`):
                 Attention dropout
             softmax_scale (`float`, *optional*):
                 The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)