class HyperConv1D(layers.Layer):
    def __init__(self, d_model, k=7, mem_size=64, hyper_dim=128, dropout=0.0):
        super().__init__()
        assert k % 2 == 1
        self.k = k
        self.d_model = d_model
        self.mem_size = mem_size

        # Input projection
        self.input_proj = layers.Dense(d_model, name="input_proj")

        # Local depthwise conv
        self.local_conv = layers.DepthwiseConv1D(kernel_size=k, padding='same', activation='silu')
        self.local_proj = layers.Dense(d_model, name="local_proj")

        # Hypernetwork: global -> scale vector
        self.hyper = tf.keras.Sequential([
            layers.Dense(hyper_dim, activation='gelu'),
            layers.Dense(d_model)
        ], name="hyper")

        # Associative memory
        self.mem_keys = self.add_weight((mem_size, d_model), initializer='glorot_uniform', trainable=True)
        self.mem_vals = self.add_weight((mem_size, d_model), initializer='glorot_uniform', trainable=True)
        self.mem_proj = layers.Dense(d_model)

        self.norm = layers.LayerNormalization()
        self.attn_pool = layers.Dense(1)

    def call(self, x):
        x_in = x
        x_dtype = x.dtype  # 입력 dtype 기억

        # 1) input projection
        x_proj = self.input_proj(x)
        # memory와 연산 위해 dtype 통일
        mem_dtype = self.mem_keys.dtype
        x_proj = tf.cast(x_proj, mem_dtype)

        # 2) local conv
        out_local = self.local_conv(x_proj)
        # hypernetwork scaling
        global_z = self.attn_pool(x_proj)
        global_z = tf.nn.softmax(global_z, axis=1)
        global_z = tf.reduce_sum(x_proj * global_z, axis=1)

        scale = tf.expand_dims(tf.nn.sigmoid(self.hyper(global_z)), 1)
        out_local = out_local * scale
        out_local = self.local_proj(out_local)


        # 3) associative memory
        sims = tf.matmul(x_proj, self.mem_keys, transpose_b=True) / tf.math.sqrt(tf.cast(self.d_model, mem_dtype))
        attn = tf.nn.softmax(sims, axis=-1)
        mem_read = tf.matmul(attn, self.mem_vals)
        mem_read = self.mem_proj(mem_read)

        # 4) fuse & residual
        out = out_local + mem_read
        out = self.norm(x_proj + out)
        out = tf.nn.silu(out)

        # 최종 출력 dtype 원래 입력 dtype으로 캐스트
        return tf.cast(out, x_dtype)