Reinstate self.scaling in Gemma3nTextAttention (#41751)

maintenance: make Gemma3nTextAttention more amenable to modular inheritance

Author: RyanMullins

src/transformers/models/gemma3n/modeling_gemma3n.py

@@ -1237,6 +1237,7 @@ def __init__(self, config: Gemma3nTextConfig, layer_idx: int):
         self.layer_idx = layer_idx
         self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
         self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = 1.0
         self.attention_dropout = self.config.attention_dropout
         self.is_causal = True
 
@@ -1335,7 +1336,7 @@ def forward(
             value_states,
             attention_mask,
             dropout=self.attention_dropout if self.training else 0.0,
-            scaling=1.0,
+            scaling=self.scaling,
             sliding_window=self.sliding_window,
             **kwargs,
         )

src/transformers/models/gemma3n/modular_gemma3n.py

@@ -1703,7 +1703,7 @@ def __init__(self, config: Gemma3nTextConfig, layer_idx: int):
         super().__init__(config, layer_idx)
         self.is_causal = True
         del self.attn_logit_softcapping
-        del self.scaling
+        self.scaling = 1.0
         self.v_norm = Gemma3nRMSNorm(dim=config.head_dim, eps=config.rms_norm_eps, with_scale=False)
 
         first_kv_shared_layer_idx = self.config.num_hidden_layers - self.config.num_kv_shared_layers
@@ -1782,7 +1782,7 @@ def forward(
             value_states,
             attention_mask,
             dropout=self.attention_dropout if self.training else 0.0,
-            scaling=1.0,
+            scaling=self.scaling,
             sliding_window=self.sliding_window,
             **kwargs,
         )