Gemma capping is a must for big models (#31698)

* softcapping

* soft cap before the mask

* style

* ...

* super nit

src/transformers/models/gemma2/configuration_gemma2.py

@@ -78,6 +78,7 @@ class Gemma2Config(PretrainedConfig):
         attention_dropout (`float`, *optional*, defaults to 0.0):
             The dropout ratio for the attention probabilities.
         final_logit_softcapping (`float`, *optional*, defaults to 30.0): scaling factor when applying tanh softcapping on the logits.
+        attn_logit_softcapping (`float`, *optional*, defaults to 50.0): scaling factor when applying tanh softcapping on the attention scores.
         query_pre_attn_scalar (`float`, *optional*, defaults to 224): scaling factor used on the attention scores
         sliding_window (`int`, *optional*, defaults to 4096): in Gemma2, every other layer uses sliding window attention. This is the
             size of the sliding window.
@@ -116,6 +117,7 @@ class Gemma2Config(PretrainedConfig):
         attention_bias=False,
         attention_dropout=0.0,
         final_logit_softcapping=30.0,
+        attn_logit_softcapping=50.0,
         query_pre_attn_scalar=224,
         sliding_window=4096,
         **kwargs,
@@ -135,6 +137,7 @@ class Gemma2Config(PretrainedConfig):
         self.rope_theta = rope_theta
         self.attention_bias = attention_bias
         self.attention_dropout = attention_dropout
+        self.attn_logit_softcapping = attn_logit_softcapping
 
         super().__init__(
             pad_token_id=pad_token_id,

src/transformers/models/gemma2/modeling_gemma2.py

@@ -256,6 +256,11 @@ class Gemma2Attention(nn.Module):
 
         attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scaling
 
+        if self.config.attn_logit_softcapping is not None:
+            attn_weights = attn_weights / self.config.attn_logit_softcapping
+            attn_weights = torch.tanh(attn_weights)
+            attn_weights = attn_weights * self.config.attn_logit_softcapping
+
         if attention_mask is not None:  # no matter the length, we just slice it
             causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
             attn_weights = attn_weights + causal_mask