From a8776b7bbd086bc37dbf9a5a6b72bec7c21a7b94 Mon Sep 17 00:00:00 2001
From: Shantanu Suryawanshi <suryawanshishantanu6@gmail.com>
Date: Tue, 7 Apr 2026 12:16:12 -0400
Subject: [PATCH] Fix: Correct cross-attention query routing in Post-LN
 TransformerDecoderLayer (#3382)

---
 python/mlx/nn/layers/transformer.py |  2 +-
 python/tests/test_nn.py             | 75 +++++++++++++++++++++++++++++
 2 files changed, 76 insertions(+), 1 deletion(-)

diff --git a/python/mlx/nn/layers/transformer.py b/python/mlx/nn/layers/transformer.py
index d856f255..d7b65ed7 100644
--- a/python/mlx/nn/layers/transformer.py
+++ b/python/mlx/nn/layers/transformer.py
@@ -228,7 +228,7 @@ class TransformerDecoderLayer(Module):
             y = self.dropout1(y)
             x = self.ln1(x + y)
 
-            y = self.cross_attention(y, memory, memory, memory_mask)
+            y = self.cross_attention(x, memory, memory, memory_mask)
             y = self.dropout2(y)
             x = self.ln2(x + y)
 
diff --git a/python/tests/test_nn.py b/python/tests/test_nn.py
index 5957184b..dfeaa9f3 100644
--- a/python/tests/test_nn.py
+++ b/python/tests/test_nn.py
@@ -2006,6 +2006,81 @@ class TestLayers(mlx_tests.MLXTestCase):
         out = attn(x, x, x)
         self.assertEqual(out.shape, x.shape)
 
+    def test_transformer_encoder_layer(self):
+        # Test norm_first=True (default)
+        layer = nn.TransformerEncoderLayer(dims=32, num_heads=4)
+        x = mx.random.normal(shape=(2, 5, 32))
+        out = layer(x, mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+        # Test norm_first=False
+        layer = nn.TransformerEncoderLayer(dims=32, num_heads=4, norm_first=False)
+        out = layer(x, mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+        # Test with causal mask
+        mask = nn.MultiHeadAttention.create_additive_causal_mask(5)
+        out = layer(x, mask=mask)
+        self.assertEqual(out.shape, x.shape)
+
+        # Test with custom mlp_dims
+        layer = nn.TransformerEncoderLayer(dims=32, num_heads=4, mlp_dims=64)
+        out = layer(x, mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+    def test_transformer_decoder_layer(self):
+        dims = 32
+        num_heads = 4
+        x = mx.random.normal(shape=(2, 5, dims))
+        memory = mx.random.normal(shape=(2, 8, dims))
+
+        # Test norm_first=True (default)
+        layer = nn.TransformerDecoderLayer(dims=dims, num_heads=num_heads)
+        out = layer(x, memory, x_mask=None, memory_mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+        # Test norm_first=False
+        layer = nn.TransformerDecoderLayer(
+            dims=dims, num_heads=num_heads, norm_first=False
+        )
+        out = layer(x, memory, x_mask=None, memory_mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+        # Test with masks
+        x_mask = nn.MultiHeadAttention.create_additive_causal_mask(5)
+        out = layer(x, memory, x_mask=x_mask, memory_mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+        # Test with custom mlp_dims
+        layer = nn.TransformerDecoderLayer(dims=dims, num_heads=num_heads, mlp_dims=64)
+        out = layer(x, memory, x_mask=None, memory_mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+    def test_transformer_encoder(self):
+        encoder = nn.TransformerEncoder(num_layers=2, dims=32, num_heads=4)
+        x = mx.random.normal(shape=(2, 5, 32))
+        out = encoder(x, mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+    def test_transformer_decoder(self):
+        decoder = nn.TransformerDecoder(num_layers=2, dims=32, num_heads=4)
+        x = mx.random.normal(shape=(2, 5, 32))
+        memory = mx.random.normal(shape=(2, 8, 32))
+        out = decoder(x, memory, x_mask=None, memory_mask=None)
+        self.assertEqual(out.shape, x.shape)
+
+    def test_transformer(self):
+        model = nn.Transformer(
+            dims=32,
+            num_heads=4,
+            num_encoder_layers=2,
+            num_decoder_layers=2,
+        )
+        src = mx.random.normal(shape=(2, 8, 32))
+        tgt = mx.random.normal(shape=(2, 5, 32))
+        out = model(src, tgt, src_mask=None, tgt_mask=None, memory_mask=None)
+        self.assertEqual(out.shape, tgt.shape)
+
 
 if __name__ == "__main__":
     mlx_tests.MLXTestRunner()