[Pallas:MGPU] Properly restore the pytree structure when unflattening unions

Previously we returned all the leaves in a flat list, which is unhelpful.

PiperOrigin-RevId: 842637237

Author: apaszke

jax/_src/pallas/mosaic_gpu/core.py

@@ -431,18 +431,17 @@ def flatten_ref_union(ref_union: AbstractRefUnion) -> tuple[_Ref, ...]:
     union_bytes = 0
     for ref_group in ref_union.refs:
       byte_offset = 0
-      for ref in jax.tree.leaves(ref_group):
+      def unflatten(ref):
+        nonlocal byte_offset
         byte_offset = align_to(byte_offset, SMEM_ALIGNMENT)
         assert isinstance(ref, state.AbstractRef) or isinstance(
             ref, pallas_core.TransformedRef
         )
         if not isinstance(ref, pallas_core.TransformedRef):
           ref = pallas_core.TransformedRef(ref, transforms=())
         transform = ExtractAliasedRef.from_transformed_ref(ref, byte_offset)
-        flat_refs.append(
-            pallas_core.TransformedRef(
-                ref_union, transforms=(transform, *ref.transforms)
-            )
+        result = pallas_core.TransformedRef(
+            ref_union, transforms=(transform, *ref.transforms)
         )
         if jnp.issubdtype(ref.dtype, jnp.integer):
           nbits = jnp.iinfo(ref.dtype).bits
@@ -457,26 +456,29 @@ def flatten_ref_union(ref_union: AbstractRefUnion) -> tuple[_Ref, ...]:
               f" {ref.dtype}{ref.shape}"
           )
         byte_offset += ref_bits // 8
+        return result
+      flat_refs.append(jax.tree.map(unflatten, ref_group))
       union_bytes = max(union_bytes, byte_offset)
     assert union_bytes == ref_union.shape[0]
   elif ref_union.memory_space == TMEM:
     union_cols = 0
     for ref_group in ref_union.refs:
       col_offset = 0
-      for ref in jax.tree.leaves(ref_group):
+      def unflatten(ref):
+        nonlocal col_offset
         col_offset = align_to(col_offset, TMEM_COL_ALIGNMENT)
         if not isinstance(ref, pallas_core.TransformedRef):
           ref = pallas_core.TransformedRef(ref, transforms=())
         ncols = ref.layout.cols_in_shape(ref.shape,
                                          dtypes.itemsize_bits(ref.dtype))
         transform = ExtractAliasedRef.from_transformed_ref(
             ref, col_offset, layout=ref.layout)
-        flat_refs.append(
-            pallas_core.TransformedRef(
-                ref_union, transforms=(transform, *ref.transforms)
-            )
+        result = pallas_core.TransformedRef(
+            ref_union, transforms=(transform, *ref.transforms)
         )
         col_offset += ncols
+        return result
+      flat_refs.append(jax.tree.map(unflatten, ref_group))
       union_cols = max(union_cols, col_offset)
     assert union_cols == ref_union.shape[1], (union_cols, ref_union.shape[1])
   else:

tests/pallas/mosaic_gpu_test.py

@@ -1907,7 +1907,7 @@ def kernel(x_ref, y_ref, o_ref):
     y = jax.lax.iota(jnp.float32, 128) * 3
     np.testing.assert_array_equal(kernel(x, y), x + y)
 
-  def test_smem_aliasing_works(self):
+  def test_smem_aliasing_works_basic(self):
     self.skip_if_wg_semantics()
 
     in_shape = (2, 256)
@@ -1938,17 +1938,16 @@ def test_smem_aliasing_works(self):
                         plgpu.SMEM(
                             (128,),
                             jnp.float32,
-                            transforms=(plgpu.TilingTransform((64,)),),
-                    ),
+                            transforms=(plgpu.TilingTransform((64,)),)),
                     ]
                 ],
             )
         ],
     )
     def kernel(x_ref, o_ref128, aliased_ref):
-      smem_ref256, _, smem_ref128 = aliased_ref
+      smem_ref256, [_, [smem_ref128]] = aliased_ref
       # Ensure that extraction via index works the same as unfolding.
-      smem_ref128_2 = aliased_ref[2]
+      smem_ref128_2 = aliased_ref[1][1][0]
       self.assertIsInstance(smem_ref128, state_types.TransformedRef)
       self.assertIsInstance(smem_ref128_2, state_types.TransformedRef)
       self.assertIs(smem_ref128.ref, smem_ref128_2.ref)
@@ -2005,7 +2004,7 @@ def test_smem_aliasing_works_with_subbyte_dtypes(self):
         ],
     )
     def kernel(x_ref, o_refi4, aliased_ref):
-      _, smem_refi8, _, smem_refi4 = aliased_ref
+      [_, smem_refi8], [_, smem_refi4] = aliased_ref
       smem_refi8[...] = x_ref[...]
       plgpu.commit_smem()
       plgpu.copy_smem_to_gmem(smem_refi4, o_refi4)
@@ -3399,7 +3398,7 @@ def test_tmem_ref_aliasing(self):
         thread_name="x",
     )
     def kernel(x_ref, y_ref, aliased_ref, smem_ref, barrier_ref):
-      tmem_128x32a, tmem_128x32b, tmem_128x64 = aliased_ref
+      [tmem_128x32a, tmem_128x32b], tmem_128x64 = aliased_ref
       plgpu.copy_gmem_to_smem(x_ref, smem_ref, barrier_ref)
       plgpu.barrier_wait(barrier_ref)
       # Test tmem_128x32 a and b
@@ -4252,7 +4251,7 @@ def kernel(a_gmem, b_gmem, out_gmem128, out_gmem64,
       plgpu.barrier_wait(tma_barrier)
       plgpu.copy_gmem_to_smem(b_gmem, b_smem, tma_barrier)
       plgpu.barrier_wait(tma_barrier)
-      acc_128, lhs_128, lhs_64, acc_64, _ = aliased_refs
+      [acc_128, lhs_128], [lhs_64, acc_64], _ = aliased_refs
 
       # Do 128x128 @ 128x128 matmul
       plgpu.async_store_tmem(lhs_128, plgpu.load(a_smem, (), layout=plgpu.Layout.TCGEN05))
@@ -4289,21 +4288,27 @@ def kernel(a_gmem, b_gmem, out_gmem128, out_gmem64,
 
     f = self.kernel(
         kernel,
-        out_shape=[jax.ShapeDtypeStruct(shape, dtype),
-                   jax.ShapeDtypeStruct(shape, dtype)],
+        out_shape=[
+            jax.ShapeDtypeStruct(shape, dtype),
+            jax.ShapeDtypeStruct(shape, dtype),
+        ],
         scratch_shapes=[
-          plgpu.SMEM(shape, dtype, transforms=transforms),  # a_smem
-          plgpu.SMEM(shape, dtype, transforms=transforms),  # b_smem
-          plgpu.SMEM(shape, dtype, transforms=transforms),  # out_smem
-          plgpu.Barrier(),  # tma_barrier
-          plgpu.Barrier(orders_tensor_core=True),  # mma_barrier
-          plgpu.RefUnion(   # aliased_refs
-            [plgpu.TMEM((128, 128), jnp.float32), # acc
-              plgpu.TMEM((128, 128), dtype, packed=True)],  # lhs
-            [plgpu.TMEM((128, 64), dtype, packed=True),  # lhs
-             plgpu.TMEM((128, 128), jnp.float32)],  # acc
-             plgpu.TMEM((128, 128), jnp.float32)  # unused
-          ),
+            plgpu.SMEM(shape, dtype, transforms=transforms),  # a_smem
+            plgpu.SMEM(shape, dtype, transforms=transforms),  # b_smem
+            plgpu.SMEM(shape, dtype, transforms=transforms),  # out_smem
+            plgpu.Barrier(),  # tma_barrier
+            plgpu.Barrier(orders_tensor_core=True),  # mma_barrier
+            plgpu.RefUnion(  # aliased_refs
+                [
+                    plgpu.TMEM((128, 128), jnp.float32),  # acc
+                    plgpu.TMEM((128, 128), dtype, packed=True),  # lhs
+                ],
+                [
+                    plgpu.TMEM((128, 64), dtype, packed=True),  # lhs
+                    plgpu.TMEM((128, 128), jnp.float32),  # acc
+                ],
+                plgpu.TMEM((128, 128), jnp.float32),  # unused
+            ),
         ],
     )
     x = jax.random.uniform(jax.random.key(0), shape=shape, dtype=dtype)