Understanding key_dim and num_heads in tf.keras.layers.MultiHeadAttention

For example, I have input with shape (1, 1000, 10) (so, src.shape wil be (1, 1000, 10)). Then:

• This works

class Model(tf.keras.Model):
        def __init__(self):
            super(Model, self).__init__()
            self.attention1 = tf.keras.layers.MultiHeadAttention(num_heads=20, key_dim=9)
            self.dense = tf.keras.layers.Dense(10, activation="softmax")

        def call(self, src):
            output = self.attention1(src, src)
            output = tf.reshape(output, [1, 10000])
            output = self.dense(output)
            return output

• And this:

class Model(tf.keras.Model):
        def __init__(self):
            super(Model, self).__init__()
            self.attention1 = tf.keras.layers.MultiHeadAttention(num_heads=123, key_dim=17)
            self.dense = tf.keras.layers.Dense(10, activation="softmax")

        def call(self, src):
            output = self.attention1(src, src)
            output = tf.reshape(output, [1, 10000])
            output = self.dense(output)
            return output

So, this layer works with whatever num_heads and key_dim but secuence length (i.e. 1000) should be divisible by num_heads. WHY? Is it a bug? For example, the same code for Pytorch doesn't work. Also, what is a key_dim then... Thanks in advance.