Allow non-shared untraced SIT-SOT

Author: ricardoV94

pytensor/link/jax/dispatch/scan.py

@@ -60,7 +60,7 @@ def scan(*outer_inputs):
             mit_mot_init,
             mit_sot_init,
             sit_sot_init,
-            op.outer_shared(outer_inputs),
+            op.outer_untraced_sit_sot(outer_inputs),
             op.outer_non_seqs(outer_inputs),
         )  # JAX `init`
 
@@ -118,7 +118,7 @@ def inner_func_outs_to_jax_outs(
         ):
             """Convert inner_scan_func outputs into format expected by JAX scan.
 
-            old_carry + (MIT-SOT_outs, SIT-SOT_outs, NIT-SOT_outs, shared_outs) -> (new_carry, ys)
+            old_carry + (MIT-SOT_outs, SIT-SOT_outs, NIT-SOT_outs, untraced_SIT-SOT_outs) -> (new_carry, ys)
             """
             (
                 i,
@@ -133,7 +133,7 @@ def inner_func_outs_to_jax_outs(
             new_mit_sot_vals = op.inner_mitsot_outs(inner_scan_outs)
             new_sit_sot = op.inner_sitsot_outs(inner_scan_outs)
             new_nit_sot = op.inner_nitsot_outs(inner_scan_outs)
-            new_shared = op.inner_shared_outs(inner_scan_outs)
+            new_shared = op.inner_untraced_sit_sot_outs(inner_scan_outs)
 
             # New carry for next step
             # Update MIT-MOT buffer at positions indicated by output taps

pytensor/link/numba/dispatch/scan.py

@@ -108,7 +108,7 @@ def numba_funcify_Scan(op: Scan, node, **kwargs):
     outer_in_mit_sot_names = op.outer_mitsot(outer_in_names)
     outer_in_sit_sot_names = op.outer_sitsot(outer_in_names)
     outer_in_nit_sot_names = op.outer_nitsot(outer_in_names)
-    outer_in_shared_names = op.outer_shared(outer_in_names)
+    outer_in_shared_names = op.outer_untraced_sit_sot(outer_in_names)
     outer_in_non_seqs_names = op.outer_non_seqs(outer_in_names)
 
     # These are all the outer-input names that have produce outputs/have output
@@ -204,11 +204,9 @@ def add_inner_in_expr(
 
     # Inner-outputs consist of:
     # mit-mot-outputs + mit-sot-outputs + sit-sot-outputs + nit-sots +
-    # shared-outputs [+ while-condition]
+    # untraced-sit-sot-outputs [+ while-condition]
     inner_output_names = [f"inner_out_{i}" for i in range(len(op.inner_outputs))]
 
-    # inner_out_shared_names = op.inner_shared_outs(inner_output_names)
-
     # The assignment statements that copy inner-outputs into the outer-outputs
     # storage
     inner_out_to_outer_in_stmts: list[str] = []

pytensor/scan/basic.py

@@ -11,6 +11,7 @@
 from pytensor.graph.op import get_test_value
 from pytensor.graph.replace import clone_replace
 from pytensor.graph.traversal import explicit_graph_inputs
+from pytensor.graph.type import HasShape
 from pytensor.graph.utils import MissingInputError, TestValueError
 from pytensor.scan.op import Scan, ScanInfo
 from pytensor.scan.utils import expand_empty, safe_new, until
@@ -706,6 +707,12 @@ def wrap_into_list(x):
     sit_sot_inner_outputs = []
     sit_sot_rightOrder = []
 
+    n_untraced_sit_sot_outs = 0
+    untraced_sit_sot_scan_inputs = []
+    untraced_sit_sot_inner_inputs = []
+    untraced_sit_sot_inner_outputs = []
+    untraced_sit_sot_rightOrder = []
+
     # go through outputs picking up time slices as needed
     for i, init_out in enumerate(outs_info):
         # Note that our convention dictates that if an output uses
@@ -741,17 +748,35 @@ def wrap_into_list(x):
             # We need now to allocate space for storing the output and copy
             # the initial state over. We do this using the expand function
             # defined in scan utils
-            sit_sot_scan_inputs.append(
-                expand_empty(
-                    shape_padleft(actual_arg),
-                    actual_n_steps,
+            if isinstance(actual_arg.type, HasShape):
+                sit_sot_scan_inputs.append(
+                    expand_empty(
+                        shape_padleft(actual_arg),
+                        actual_n_steps,
+                    )
                 )
-            )
+                sit_sot_inner_slices.append(actual_arg)
+
+                sit_sot_inner_inputs.append(arg)
+                sit_sot_rightOrder.append(i)
+                n_sit_sot += 1
+            else:
+                # Assume variables without shape cannot be stacked (e.g., RNG variables)
+                # Because this is new, issue a warning to inform the user, except for RNG, which were the main reason for this feature
+                from pytensor.tensor.random.type import RandomType
 
-            sit_sot_inner_slices.append(actual_arg)
-            sit_sot_inner_inputs.append(arg)
-            sit_sot_rightOrder.append(i)
-            n_sit_sot += 1
+                if not isinstance(arg.type, RandomType):
+                    warnings.warn(
+                        (
+                            f"Output {actual_arg} (index {i}) with type {actual_arg.type} will be treated as untraced variable in scan. "
+                            "Only the last value will be returned, not the entire sequence."
+                        ),
+                        UserWarning,
+                    )
+                untraced_sit_sot_scan_inputs.append(actual_arg)
+                untraced_sit_sot_inner_inputs.append(arg)
+                n_untraced_sit_sot_outs += 1
+                untraced_sit_sot_rightOrder.append(i)
 
         elif init_out.get("taps", None):
             if np.any(np.array(init_out.get("taps", [])) > 0):
@@ -802,9 +827,10 @@ def wrap_into_list(x):
         #      a map); in that case we do not have to do anything ..
 
     # Re-order args
-    max_mit_sot = np.max([-1, *mit_sot_rightOrder]) + 1
-    max_sit_sot = np.max([-1, *sit_sot_rightOrder]) + 1
-    n_elems = np.max([max_mit_sot, max_sit_sot])
+    max_mit_sot = max(mit_sot_rightOrder, default=-1) + 1
+    max_sit_sot = max(sit_sot_rightOrder, default=-1) + 1
+    max_untraced_sit_sot_outs = max(untraced_sit_sot_rightOrder, default=-1) + 1
+    n_elems = np.max((max_mit_sot, max_sit_sot, max_untraced_sit_sot_outs))
     _ordered_args = [[] for x in range(n_elems)]
     offset = 0
     for idx in range(n_mit_sot):
@@ -825,6 +851,11 @@ def wrap_into_list(x):
         else:
             _ordered_args[sit_sot_rightOrder[idx]] = [sit_sot_inner_inputs[idx]]
 
+    for idx in range(n_untraced_sit_sot_outs):
+        _ordered_args[untraced_sit_sot_rightOrder[idx]] = [
+            untraced_sit_sot_inner_inputs[idx]
+        ]
+
     ordered_args = list(chain.from_iterable(_ordered_args))
     if single_step_requested:
         args = inner_slices + ordered_args + non_seqs
@@ -842,6 +873,11 @@ def wrap_into_list(x):
         raw_inner_outputs = fn(*args)
 
     condition, outputs, updates = get_updates_and_outputs(raw_inner_outputs)
+    if updates:
+        warnings.warn(
+            "Updates functionality in Scan are deprecated. Use explicit outputs_info and build shared update expressions manually, even for RNGs.",
+            DeprecationWarning,  # Only meant for developers for now, not users. Switch to FutureWarning later, before removing.
+        )
     if condition is not None:
         as_while = True
     else:
@@ -883,6 +919,8 @@ def wrap_into_list(x):
     fake_outputs = clone_replace(
         outputs, replace=dict(zip(non_seqs, fake_nonseqs, strict=True))
     )
+    # TODO: Once we don't treat shared variables specially we should use `truncated_graph_inputs`
+    #  to find implicit inputs in a way that reduces the size of the inner function
     known_inputs = [*args, *fake_nonseqs]
     extra_inputs = [
         x for x in explicit_graph_inputs(fake_outputs) if x not in known_inputs
@@ -939,18 +977,19 @@ def wrap_into_list(x):
         if "taps" in out and out["taps"] != [-1]:
             mit_sot_inner_outputs.append(outputs[i])
 
-    # Step 5.2 Outputs with tap equal to -1
+    # Step 5.2 Outputs with tap equal to -1 (traced and untraced)
     for i, out in enumerate(outs_info):
         if "taps" in out and out["taps"] == [-1]:
-            sit_sot_inner_outputs.append(outputs[i])
+            output = outputs[i]
+            if isinstance(output.type, HasShape):
+                sit_sot_inner_outputs.append(output)
+            else:
+                untraced_sit_sot_inner_outputs.append(output)
 
     # Step 5.3 Outputs that correspond to update rules of shared variables
-    inner_replacements = {}
-    n_shared_outs = 0
-    shared_scan_inputs = []
-    shared_inner_inputs = []
-    shared_inner_outputs = []
+    # This whole special logic for shared variables is deprecated
     sit_sot_shared = []
+    inner_replacements = {}
     no_update_shared_inputs = []
     for input in dummy_inputs:
         if not isinstance(input.variable, SharedVariable):
@@ -1003,10 +1042,10 @@ def wrap_into_list(x):
                 sit_sot_shared.append(input.variable)
 
             else:
-                shared_inner_inputs.append(new_var)
-                shared_scan_inputs.append(input.variable)
-                shared_inner_outputs.append(input.update)
-                n_shared_outs += 1
+                untraced_sit_sot_inner_inputs.append(new_var)
+                untraced_sit_sot_scan_inputs.append(input.variable)
+                untraced_sit_sot_inner_outputs.append(input.update)
+                n_untraced_sit_sot_outs += 1
         else:
             no_update_shared_inputs.append(input)
 
@@ -1071,7 +1110,7 @@ def wrap_into_list(x):
         + mit_mot_inner_inputs
         + mit_sot_inner_inputs
         + sit_sot_inner_inputs
-        + shared_inner_inputs
+        + untraced_sit_sot_inner_inputs
         + other_shared_inner_args
         + other_inner_args
     )
@@ -1081,7 +1120,7 @@ def wrap_into_list(x):
         + mit_sot_inner_outputs
         + sit_sot_inner_outputs
         + nit_sot_inner_outputs
-        + shared_inner_outputs
+        + untraced_sit_sot_inner_outputs
     )
     if condition is not None:
         inner_outs.append(condition)
@@ -1101,7 +1140,7 @@ def wrap_into_list(x):
         mit_mot_out_slices=tuple(tuple(v) for v in mit_mot_out_slices),
         mit_sot_in_slices=tuple(tuple(v) for v in mit_sot_tap_array),
         sit_sot_in_slices=tuple((-1,) for x in range(n_sit_sot)),
-        n_shared_outs=n_shared_outs,
+        n_untraced_sit_sot_outs=n_untraced_sit_sot_outs,
         n_nit_sot=n_nit_sot,
         n_non_seqs=len(other_shared_inner_args) + len(other_inner_args),
         as_while=as_while,
@@ -1127,7 +1166,7 @@ def wrap_into_list(x):
         + mit_mot_scan_inputs
         + mit_sot_scan_inputs
         + sit_sot_scan_inputs
-        + shared_scan_inputs
+        + untraced_sit_sot_scan_inputs
         + [actual_n_steps for x in range(n_nit_sot)]
         + other_shared_scan_args
         + other_scan_args
@@ -1173,13 +1212,28 @@ def remove_dimensions(outs, offsets=None):
     nit_sot_outs = remove_dimensions(scan_outs[offset : offset + n_nit_sot])
 
     offset += n_nit_sot
-    for idx, update_rule in enumerate(scan_outs[offset : offset + n_shared_outs]):
-        update_map[shared_scan_inputs[idx]] = update_rule
 
-    _scan_out_list = mit_sot_outs + sit_sot_outs + nit_sot_outs
+    # Legacy support for explicit untraced sit_sot and those built with update dictionary
+    # Switch to n_untraced_sit_sot_outs after deprecation period
+    n_explicit_untraced_sit_sot_outs = len(untraced_sit_sot_rightOrder)
+    untraced_sit_sot_outs = scan_outs[
+        offset : offset + n_explicit_untraced_sit_sot_outs
+    ]
+
+    # Legacy support: map shared outputs to their updates
+    offset += n_explicit_untraced_sit_sot_outs
+    for idx, update_rule in enumerate(scan_outs[offset:]):
+        update_map[untraced_sit_sot_scan_inputs[idx]] = update_rule
+
+    _scan_out_list = mit_sot_outs + sit_sot_outs + nit_sot_outs + untraced_sit_sot_outs
     # Step 10. I need to reorder the outputs to be in the order expected by
     # the user
-    rightOrder = mit_sot_rightOrder + sit_sot_rightOrder + nit_sot_rightOrder
+    rightOrder = (
+        mit_sot_rightOrder
+        + sit_sot_rightOrder
+        + untraced_sit_sot_rightOrder
+        + nit_sot_rightOrder
+    )
     scan_out_list = [None] * len(rightOrder)
     for idx, pos in enumerate(rightOrder):
         if pos >= 0: