From 34778daf3e7b6a9b72e58fbb999671c0da1fe015 Mon Sep 17 00:00:00 2001
From: David Hagen <david@drhagen.com>
Date: Thu, 13 Aug 2026 06:59:07 -0400
Subject: [PATCH] Shrink over-allocated sparse layers to final size

---
 .../iteration_graph/outputs/_append.py        | 54 ++++++++++++++-----
 tests/test_combinatorically.py                | 15 ++++++
 tests_cffi/test_combinatorically.py           | 15 ++++++
 3 files changed, 72 insertions(+), 12 deletions(-)

diff --git a/src/tensora/iteration_graph/outputs/_append.py b/src/tensora/iteration_graph/outputs/_append.py
index 3b0df54..1c588b0 100644
--- a/src/tensora/iteration_graph/outputs/_append.py
+++ b/src/tensora/iteration_graph/outputs/_append.py
@@ -4,7 +4,14 @@
 
 from ...format import Mode
 from ...ir import SourceBuilder, types
-from ...ir.ast import ArrayAllocate, Expression, IntegerLiteral, Multiply, Variable
+from ...ir.ast import (
+    ArrayAllocate,
+    ArrayReallocate,
+    Expression,
+    IntegerLiteral,
+    Multiply,
+    Variable,
+)
 from ...kernel_type import KernelType
 from .._names import (
     crd_capacity_name,
@@ -104,24 +111,47 @@ def write_cleanup(self, kernel_type: KernelType):
             target_name = self.output.name
             output_tensor = Variable(target_name)
 
+            # The number of positions in the previous layer; each dense layer multiplies it by its
+            # dimension, while each compressed layer replaces it with its final layer pointer.
+            previous_size: Expression = IntegerLiteral(1)
+            all_dense = True
             for i, mode in enumerate(self.output.modes):
                 if mode == Mode.dense:
-                    pass
+                    previous_size = previous_size.times(dimension_name(self.output.indexes[i]))
                 elif mode == Mode.compressed:
+                    pos_array = pos_name(target_name, i)
+                    if not all_dense:
+                        # If any previous layer was compressed, pos was allocated with a guessed
+                        # capacity, so shrink it to its final size
+                        source.append(
+                            pos_array.assign(
+                                ArrayReallocate(pos_array, types.integer, previous_size.plus(1))
+                            )
+                        )
+
+                    # crd is always allocated with a guessed capacity, so shrink it to its final
+                    # size
+                    crd_array = crd_name(target_name, i)
+                    final_size = layer_pointer(self.output.id, i)
                     source.append(
-                        output_tensor.attr("indices")
-                        .idx(i)
-                        .idx(0)
-                        .assign(pos_name(target_name, i))
-                    )
-                    source.append(
-                        output_tensor.attr("indices")
-                        .idx(i)
-                        .idx(1)
-                        .assign(crd_name(target_name, i))
+                        crd_array.assign(ArrayReallocate(crd_array, types.integer, final_size))
                     )
+
+                    source.append(output_tensor.attr("indices").idx(i).idx(0).assign(pos_array))
+                    source.append(output_tensor.attr("indices").idx(i).idx(1).assign(crd_array))
+
+                    previous_size = final_size
+                    all_dense = False
                 else:
                     raise NotImplementedError()
+
+            if not all_dense:
+                # If any layer was compressed, vals was allocated with a guessed capacity, so
+                # shrink it to its final size
+                vals_array = vals_name(target_name)
+                source.append(
+                    vals_array.assign(ArrayReallocate(vals_array, types.float, previous_size))
+                )
             source.append(output_tensor.attr("vals").assign(vals_name(target_name)))
 
         return source
diff --git a/tests/test_combinatorically.py b/tests/test_combinatorically.py
index 242e95e..a805113 100644
--- a/tests/test_combinatorically.py
+++ b/tests/test_combinatorically.py
@@ -32,6 +32,21 @@ def test_copy_2_backwards(dense, format_in, format_out):
     assert actual == a
 
 
+@pytest.mark.parametrize(
+    "dense",
+    [
+        [[[0, 2, 4], [0, -1, 0]], [[0, 0, 3], [5, 0, 0]]],
+        [[[0, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]]],
+    ],
+)
+@pytest.mark.parametrize("format_in", ["ddd", "dds", "dsd", "dss", "sdd", "sds", "ssd", "sss"])
+@pytest.mark.parametrize("format_out", ["ddd", "dds", "dsd", "dss", "sdd", "sds", "ssd", "sss"])
+def test_copy_3(dense, format_in, format_out):
+    a = Tensor.from_lol(dense, format=format_in)
+    actual = evaluate("b(i,j,k) = a(i,j,k)", format_out, a=a)
+    assert actual == a
+
+
 @pytest.mark.parametrize("expression", [0, 1])
 def test_constant_scalar(expression):
     actual = evaluate(f"a() = {expression}", "")
diff --git a/tests_cffi/test_combinatorically.py b/tests_cffi/test_combinatorically.py
index 60c21be..fda618e 100644
--- a/tests_cffi/test_combinatorically.py
+++ b/tests_cffi/test_combinatorically.py
@@ -43,6 +43,21 @@ def test_copy_2_backwards(dense, format_in, format_out):
     assert actual == a
 
 
+@pytest.mark.parametrize(
+    "dense",
+    [
+        [[[0, 2, 4], [0, -1, 0]], [[0, 0, 3], [5, 0, 0]]],
+        [[[0, 0, 0], [0, 0, 0]], [[0, 0, 0], [0, 0, 0]]],
+    ],
+)
+@pytest.mark.parametrize("format_in", ["ddd", "dds", "dsd", "dss", "sdd", "sds", "ssd", "sss"])
+@pytest.mark.parametrize("format_out", ["ddd", "dds", "dsd", "dss", "sdd", "sds", "ssd", "sss"])
+def test_copy_3(dense, format_in, format_out):
+    a = Tensor.from_lol(dense, format=format_in)
+    actual = evaluate("b(i,j,k) = a(i,j,k)", format_out, a=a)
+    assert actual == a
+
+
 @pytest.mark.parametrize("expression", [0, 1])
 def test_constant_scalar(expression):
     actual = evaluate(f"a() = {expression}", "")