Update Mo.py

Mo.py

@@ -125,39 +125,59 @@ class SwiGLU(layers.Layer):
         return self.out(x_val * tf.nn.silu(x_gate))
 
 class LoU(layers.Layer):
-    def __init__(self, d_model, clip_value=5.0, eps=1e-6):
-        super().__init__()
-        self.d_model = d_model
-        self.clip_value = float(clip_value)
-        self.eps = float(eps)
-        self.Q = layers.Dense(d_model, dtype='float32')
-        self.K = layers.Dense(d_model, dtype='float32')
-        self.V = layers.Dense(d_model, dtype='float32')
-        self.norm = layers.LayerNormalization(epsilon=1e-5, dtype='float32')
-        self.norm1 = layers.LayerNormalization(epsilon=1e-5, dtype='float32') 
-        self.glu = SwiGLU(d_model, 320)
-    def call(self, x):
-        x_f32 = tf.cast(x, tf.float32)
-        residual = x_f32
-        x_f32 = self.norm1(x)
-        q = self.Q(x_f32)
-        k = self.K(x_f32)
-        V = self.V(x_f32)
-        g_q = (tf.nn.tanh(q) + 1.0) / 2.0
-        g_k = (tf.nn.tanh(k) + 1.0) / 2.0
-        score = g_q * g_k
-        score = tf.cumsum(score, axis=1)
-        seq_len = tf.shape(score)[1]
-        count_for_mean = tf.cast(tf.range(seq_len) + 1, score.dtype)
-        count_for_mean = tf.reshape(count_for_mean, (1, seq_len, 1))
-        score_mean = score / count_for_mean
-        denom = tf.maximum(score_mean, self.eps)
-        score_norm = score / denom
-        score_clipped = tf.clip_by_value(score_norm, -self.clip_value, self.clip_value)
-        x_comb = score_clipped * V
-        out = self.norm(x_comb + residual)
-        out = self.glu(out)
-        return tf.cast(out, x.dtype)
+    def __init__(self, d_model, clip_value=5.0, eps=1e-6, dropout_rate=0.1):
+        super().__init__()
+        self.d_model = d_model
+        self.clip_value = float(clip_value)
+        self.eps = float(eps)
+        self.dropout_rate = dropout_rate
+        
+        self.Q = layers.Dense(d_model, dtype='float32')
+        self.K = layers.Dense(d_model, dtype='float32')
+        self.V = layers.Dense(d_model, dtype='float32')
+        
+        self.norm = layers.LayerNormalization(epsilon=1e-5, dtype='float32')
+        self.norm1 = layers.LayerNormalization(epsilon=1e-5, dtype='float32')  
+        self.glu = SwiGLU(d_model, 320)
+        self.dropout = layers.Dropout(dropout_rate)
+
+    def call(self, x, training=False):
+        x_f32 = tf.cast(x, tf.float32)
+        residual = x_f32
+        
+        x_f32 = self.norm1(x_f32)
+        q = self.Q(x_f32)
+        k = self.K(x_f32)
+        V = self.V(x_f32)
+        
+        # gating
+        g_q = (tf.nn.tanh(q) + 1.0) / 2.0
+        g_k = (tf.nn.tanh(k) + 1.0) / 2.0
+        
+        # cumulative score
+        score = g_q * g_k
+        score = tf.cumsum(score, axis=1)
+        
+        seq_len = tf.shape(score)[1]
+        count_for_mean = tf.cast(tf.range(seq_len) + 1, score.dtype)
+        count_for_mean = tf.reshape(count_for_mean, (1, seq_len, 1))
+        score_mean = score / count_for_mean
+        
+        # normalization + softmax-ish
+        denom = tf.maximum(score_mean, self.eps)
+        score_norm = score / denom
+        score_norm = tf.nn.softmax(score_norm, axis=1)
+        
+        # clipping + dropout
+        score_clipped = tf.clip_by_value(score_norm, -self.clip_value, self.clip_value)
+        score_clipped = self.dropout(score_clipped, training=training)
+        
+        x_comb = score_clipped * V
+        out = self.norm(x_comb + residual)
+        out = self.glu(out)
+        
+        return tf.cast(out, x.dtype)
+
 
 class Lo(layers.Layer):
     def __init__(self, d_model):