Avoid large allocation for taps of length 1 in ScanSaveMem

Author: ricardoV94

pytensor/scan/rewriting.py

@@ -53,6 +53,7 @@
 from pytensor.tensor.basic import (
     Alloc,
     AllocEmpty,
+    atleast_Nd,
     get_scalar_constant_value,
 )
 from pytensor.tensor.elemwise import DimShuffle, Elemwise
@@ -1186,8 +1187,8 @@ def while_scan_merge_subtensor_last_element(fgraph, scan_node):
     return subtensor_merge_replacements
 
 
-def _is_default_scan_buffer(x: TensorVariable) -> bool:
-    node = x.owner
+def _is_default_scan_buffer(final_buffer: TensorVariable, taps: int) -> bool:
+    node = final_buffer.owner
 
     if node is None:
         return False
@@ -1200,8 +1201,10 @@ def _is_default_scan_buffer(x: TensorVariable) -> bool:
     ):
         return False
 
-    x, y, *_ = node.inputs
-    if not (x.owner is not None and isinstance(x.owner.op, AllocEmpty)):
+    init_buffer, init_value, *_ = node.inputs
+    if not (
+        init_buffer.owner is not None and isinstance(init_buffer.owner.op, AllocEmpty)
+    ):
         return False
 
     # The value may have been broadcast to fill in the initial taps.
@@ -1218,10 +1221,16 @@ def _is_default_scan_buffer(x: TensorVariable) -> bool:
     #   1. alloc_empty(2 + nsteps)[:2].broadcastable == x.broadcastable
     # But due to laziness we use the slightly more conservative check:
     #   2. alloc_empty(2 + nsteps).broadcastable == x.broadcastable
-    if broadcasted_by(y, x):
-        return False
-
-    return True
+    if taps > 1:
+        return not broadcasted_by(init_value, init_buffer)
+    else:
+        # In this case we know we have alloc_empty(1 + nsteps, ...)[:1].set(init_value)
+        # The first dimension cannot possibly broadcast in the subtensor assignment,
+        # so we exclude it from `broadcasted_by`. To exclude it we squeeze it out,
+        # after adding any other implicit expand_dims. We select into the first entry of
+        # the buffer, to check for potential broadcasting in other dimensions.
+        init_value_ = atleast_Nd(init_value, n=init_buffer.ndim)
+        return not broadcasted_by(init_value_.squeeze(0), init_buffer[0])
 
 
 def scan_save_mem_rewrite(fgraph, node, backend_supports_output_pre_allocation: bool):
@@ -1574,15 +1583,16 @@ def scan_save_mem_rewrite(fgraph, node, backend_supports_output_pre_allocation:
                 # If the memory for this output has been pre-allocated
                 # before going into the scan op (by an alloc node)
                 if idx < op_info.n_mit_sot + op_info.n_sit_sot:
+                    taps = init_l[i]
                     nw_input = nw_inputs[offset + idx]
 
                     # Recreate default buffers with new size
-                    if _is_default_scan_buffer(nw_input):
-                        extra_size = 1 if required_orphan else val - init_l[i]
+                    if _is_default_scan_buffer(nw_input, taps):
+                        extra_size = 1 if required_orphan else val - taps
                         nw_input = expand_empty(nw_input.owner.inputs[1], extra_size)
                     # Otherwise, just trim with a slice
                     else:
-                        stop = init_l[i] if required_orphan else val
+                        stop = taps if required_orphan else val
                         nw_input = nw_input[:stop]
 
                     nw_inputs[offset + idx] = nw_input
@@ -1626,14 +1636,13 @@ def scan_save_mem_rewrite(fgraph, node, backend_supports_output_pre_allocation:
                     # val == 0 means that we want to keep all intermediate
                     # results for that state, including the initial values.
                     if idx < op_info.n_mit_sot + op_info.n_sit_sot:
+                        taps = init_l[op_info.n_mit_mot + idx]
                         in_idx = offset + idx
                         nw_input = nw_inputs[in_idx]
-                        if _is_default_scan_buffer(nw_input):
+                        if _is_default_scan_buffer(nw_input, taps):
                             nw_input = expand_empty(nw_input.owner.inputs[1], nw_steps)
                         else:
-                            # Number of steps in the initial state
-                            init_l_pt = pt.as_tensor(init_l[op_info.n_mit_mot + idx])
-                            nw_input = nw_input[: (init_l_pt + nw_steps)]
+                            nw_input = nw_input[: (taps + nw_steps)]
                         nw_inputs[in_idx] = nw_input
 
                     elif (

pytensor/tensor/rewriting/basic.py

@@ -95,9 +95,11 @@ def broadcasted_by(x: TensorVariable, y: TensorVariable) -> bool:
     """
     bx = x.type.broadcastable
     by = y.type.broadcastable
-    if len(bx) < len(by):
+    bx_len = len(bx)
+    by_len = len(by)
+    if bx_len < by_len:
         return True
-    bx = bx[-len(by) :]
+    bx = bx[bx_len - by_len :]
     return any(bx_dim and not by_dim for bx_dim, by_dim in zip(bx, by, strict=True))
 
 

tests/scan/test_rewriting.py

@@ -9,13 +9,14 @@
 from pytensor.compile.mode import get_default_mode
 from pytensor.configdefaults import config
 from pytensor.gradient import grad, jacobian
-from pytensor.graph.basic import Constant, equal_computations
+from pytensor.graph.basic import Constant, ancestors, equal_computations
 from pytensor.graph.fg import FunctionGraph
 from pytensor.graph.replace import clone_replace
 from pytensor.scan.op import Scan
 from pytensor.scan.rewriting import ScanInplaceOptimizer, ScanMerge
 from pytensor.scan.utils import until
 from pytensor.tensor import stack
+from pytensor.tensor.basic import AllocEmpty
 from pytensor.tensor.blas import Dot22
 from pytensor.tensor.elemwise import Elemwise
 from pytensor.tensor.math import Dot, dot, sigmoid, tanh
@@ -1207,7 +1208,7 @@ def test_inplace3(self):
 
 
 class TestSaveMem:
-    mode = get_default_mode().including("scan_save_mem")
+    mode = get_default_mode().including("scan_save_mem").excluding("scan_pushout")
 
     def test_save_mem(self):
         rng = np.random.default_rng(utt.fetch_seed())
@@ -1371,7 +1372,7 @@ def test_save_mem_cannot_reduce_constant_number_of_steps(self):
         )
 
     def test_save_mem_store_steps(self):
-        def f_rnn(u_t, x1_tm1, x1_tm3, x2_tm1, x3tm2, x3_tm1, x4_tm1):
+        def step(u_t, x1_tm1, x1_tm3, x2_tm1, x3tm2, x3_tm1, x4_tm1):
             return (
                 u_t + 1.0,
                 u_t + 2.0,
@@ -1388,7 +1389,7 @@ def f_rnn(u_t, x1_tm1, x1_tm3, x2_tm1, x3tm2, x3_tm1, x4_tm1):
         x30 = vector("x30")
         x40 = scalar("x40")
         [x1, x2, x3, x4, x5, x6, x7], updates = scan(
-            f_rnn,
+            step,
             u,
             [
                 None,
@@ -1404,7 +1405,7 @@ def f_rnn(u_t, x1_tm1, x1_tm3, x2_tm1, x3tm2, x3_tm1, x4_tm1):
             go_backwards=False,
         )
 
-        f2 = function(
+        f = function(
             [u, x10, x20, x30, x40],
             [x1[-7], x2[-3:-1], x3[-6:], x4[-1], x5[-1]],
             updates=updates,
@@ -1417,13 +1418,51 @@ def f_rnn(u_t, x1_tm1, x1_tm3, x2_tm1, x3tm2, x3_tm1, x4_tm1):
         v_u = rng.uniform(-5.0, 5.0, size=(20,))
 
         # compute the output in numpy
-        tx1, tx2, tx3, tx4, tx5 = f2(v_u, [0, 0], 0, [0, 0], 0)
-
-        utt.assert_allclose(tx1, v_u[-7] + 1.0)
-        utt.assert_allclose(tx2, v_u[-3:-1] + 2.0)
-        utt.assert_allclose(tx3, v_u[-6:] + 3.0)
-        utt.assert_allclose(tx4, v_u[-1] + 4.0)
-        utt.assert_allclose(tx5, v_u[-1] + 5.0)
+        tx1, tx2, tx3, tx4, tx5 = f(v_u, [0, 0], 0, [0, 0], 0)
+        rtol = 1e-7 if config.floatX == "float64" else 1e-6
+        np.testing.assert_allclose(tx1, v_u[-7] + 1.0, rtol=rtol)
+        np.testing.assert_allclose(tx2, v_u[-3:-1] + 2.0, rtol=rtol)
+        np.testing.assert_allclose(tx3, v_u[-6:] + 3.0, rtol=rtol)
+        np.testing.assert_allclose(tx4, v_u[-1] + 4.0, rtol=rtol)
+        np.testing.assert_allclose(tx5, v_u[-1] + 5.0, rtol=rtol)
+
+        # Confirm reduction in buffer sizes
+        [scan_node] = [
+            node for node in f.maker.fgraph.apply_nodes if isinstance(node.op, Scan)
+        ]
+        # x6 and x7 are dropped because they are not used
+        [n_steps, seq, x4_buffer, x5_buffer, x1_len, x2_len, x3_len] = scan_node.inputs
+        [x4_underlying_alloc] = [
+            var
+            for var in ancestors([x4_buffer])
+            if var.owner and isinstance(var.owner.op, AllocEmpty)
+        ]
+        [x5_underlying_alloc] = [
+            var
+            for var in ancestors([x5_buffer])
+            if var.owner and isinstance(var.owner.op, AllocEmpty)
+        ]
+        buffer_lengths = pytensor.function(
+            [u, x10, x20, x30, x40],
+            [
+                x1_len,
+                x2_len,
+                x3_len,
+                x4_underlying_alloc.shape[0],
+                x5_underlying_alloc.shape[0],
+            ],
+            accept_inplace=True,
+            on_unused_input="ignore",
+            allow_input_downcast=True,
+        )(v_u, [0, 0], 0, [0, 0], 0)
+        # ScanSaveMem keeps +1 entries to handle taps with preallocated outputs
+        assert [int(i) for i in buffer_lengths] == [
+            7,  # entry -7 of a map variable is kept, we need at least that many
+            3,  # entries [-3, -2] of a map variable are kept, we need at least 3
+            6,  # last six entries of a map variable are kept
+            2 + 1,  # last entry of a double tap variable is kept
+            1 + 1,  # last entry of a single tap variable is kept
+        ]
 
     def test_savemem_does_not_duplicate_number_of_scan_nodes(self):
         var = pt.ones(())