[JAX] Fixed the shape miss-matching issue in MLP. (#859)

* Fixed the shape mismatching issue in MLP.
Signed-off-by: Ming Huang <mingh@nvidia.com>

* Add a corresponding test
Signed-off-by: Ming Huang <mingh@nvidia.com>

---------
Signed-off-by: Ming Huang <mingh@nvidia.com>
Co-authored-by: Phuong Nguyen <36155692+phu0ngng@users.noreply.github.com>

tests/jax/test_layer.py

@@ -177,6 +177,8 @@ ATTRS = [{}, {
     _KEY_OF_SELF_ATTN_BIAS_TYPE: "no_bias",
 }, {
     _KEY_OF_ATTENTION_DROPOUT: 0.3,
+}, {
+    _KEY_OF_MLP_ACTIVATIONS: (('relu', 'relu')),
 }]
 
 ATTRS = [{**BASE_ATTRS, **attr} for attr in ATTRS]

transformer_engine/jax/flax/module.py

@@ -1148,8 +1148,8 @@ class LayerNormMLP(TransformerEngineBase):
                     x_i = _convert_to_activation_function(act_fn)(x[idx])
                     activations.append(x_i)
                 z = functools.reduce(operator.mul, activations)
-                if num_activations == 1:
-                    z = jnp.reshape(z, (*z.shape[:-2], -1))
+                # Remove act axis
+                z = jnp.reshape(z, (*z.shape[:-2], -1))
 
             z = nn.Dropout(rate=self.intermediate_dropout_rate,
                            broadcast_dims=self.intermediate_hidden_dropout_dims,