[SOT][dynamic shape] add symbolic fallback when SymbolicVariable's handler not exists.

python/paddle/jit/sot/infer_meta.py

@@ -310,10 +310,9 @@ def infer_meta(self, func, *args, **kwargs):
                 if isinstance(func, str):
                     # TODO(Aurelius84): Is length of args always greater than 0?
                     # Do we need add condition check here?
-                    out = getattr(args[0], func)(*args[1:], **kwargs)
-                else:
-                    out = func(*args, **kwargs)
-
+                    func = getattr(args[0], func)
+                    args = args[1:]
+                out = func(*args, **kwargs)
         return convert_variable_to_meta_info(out)
 
 

python/paddle/jit/sot/opcode_translator/executor/executor_cache.py

@@ -65,7 +65,9 @@ def __init__(self):
         self.translate_count = 0
         self.code_symbolic_inputs = {}
 
-    def get_symbolic_inputs(self, code: types.CodeType):
+    def get_symbolic_inputs(
+        self, code: types.CodeType
+    ) -> dict[str, dict[int, int] | None]:
         self.code_symbolic_inputs.setdefault(code, {})
         return self.code_symbolic_inputs[code]
 

python/paddle/jit/sot/opcode_translator/executor/variable_dispatch.py

@@ -640,13 +640,14 @@ def dispatch_reversed(var: ContainerVariable):
 # bool
 Dispatcher.register(
     bool,
-    ("ContainerVariable | SymbolicVariable",),
+    ("ContainerVariable",),
     lambda var: var.bool(),
 )
+
 Dispatcher.register(
     operator.truth,
-    ("ConstantVariable | SymbolicVariable",),
-    lambda var: var.bool(),
+    ("ConstantVariable",),
+    lambda var: Dispatcher.call(bool, var),
 )
 
 # str
@@ -936,7 +937,7 @@ def is_not_func(var: VariableBase, other: VariableBase):
                 binary_fn,
             ),
         )
-# Tensor
+# Tensor and Symbolic
 fallback_tensor_unary_method = {
     int,
     bool,
@@ -1034,20 +1035,6 @@ def tensor_mod_dispatcher(
                         magic_method.name,
                     ),
                 )
-# Symbolic
-for unary_fn in fallback_tensor_unary_method:
-    Dispatcher.register(
-        unary_fn,
-        ("SymbolicVariable",),
-        partial(
-            lambda fn, var: VariableFactory.from_value(
-                fn(var.get_py_value()),
-                var.graph,
-                tracker=DummyTracker([var]),
-            ),
-            unary_fn,
-        ),
-    )
 
 for binary_fn in BINARY_OPS:
     for magic_method in magic_method_builtin_dispatch(binary_fn):

python/paddle/jit/sot/opcode_translator/executor/variables/basic.py

@@ -363,9 +363,9 @@ def analyse_dynamic_axes(self, tracker: Tracker):
                 1,
                 f"Start analyse dynamic axes for {tracker.trace_value_from_frame().inlined_expr} in {self.graph.pycode_gen._origin_code}\n",
             )
-            for key in symbolic_inputs:
+            for key, symbolic_input in symbolic_inputs.items():
                 if key.startswith(tracker_expr):
-                    log(1, f"  {key}: {symbolic_inputs[key]}\n")
+                    log(1, f"  {key}: {symbolic_input}\n")
             log(
                 1,
                 f"  -> Tensor {tracker_expr} with dynamic axes {dynamic_axes}\n",
@@ -666,6 +666,7 @@ class SymbolicVariable(VariableBase):
 
     var_name_generator = NameGenerator("symint_")
     value: int | SymbolicValue
+    mutable_attrs = ["need_guard_value"]
 
     def __init__(
         self,
@@ -685,6 +686,12 @@ def __init__(
                 [], paddle.int64, True, self.var_name, False, None, None
             )
         self.need_guard_value = False
+        self.graph.side_effects.record_mutable_variable(self)
+
+    def to_constant(self):
+        return ConstantVariable(
+            self.get_py_value(), self.graph, DummyTracker([self])
+        )
 
     def get_py_value(self, allow_tensor: bool = False) -> bool | int | float:
         self.need_guard_value = True
@@ -748,11 +755,6 @@ def make_stringified_guard(self) -> list[StringifiedExpression]:
 
         assert frame_value_tracer.inlined_expr in symbolic_inputs
 
-        # TODO(zrr1999): Once dynamic shape is used, there will be no new guards
-        if isinstance(self.value, int):
-            symbolic_input = symbolic_inputs[frame_value_tracer.inlined_expr]
-            symbolic_input.setdefault(self.value, 0)
-            symbolic_input[self.value] += 1
         if self.need_guard_value:
             return super().make_stringified_guard()
         return [
@@ -765,12 +767,16 @@ def make_stringified_guard(self) -> list[StringifiedExpression]:
 
     @staticmethod
     def should_create_symbolic_variable(
-        value: Any, tracker: Tracker, symbolic_inputs: dict[str, dict[int, int]]
+        value: Any,
+        tracker: Tracker,
+        symbolic_inputs: dict[str, dict[int, int] | None],
     ):
         tracker_expr = tracker.trace_value_from_frame().inlined_expr
         symbolic_inputs.setdefault(tracker_expr, {})
         if tracker_expr in symbolic_inputs:
             symbolic_input = symbolic_inputs[tracker_expr]
+            if symbolic_input is None:
+                return False
             symbolic_input.setdefault(value, 0)
             symbolic_input[value] += 1
             if symbolic_input[value] >= STATIC_DIM_FREQ_THRESHOLD:

python/paddle/jit/sot/opcode_translator/executor/variables/callable.py

@@ -15,16 +15,22 @@
 from __future__ import annotations
 
 import inspect
+import itertools
 import operator
 import types
 from functools import reduce
-from typing import TYPE_CHECKING, Any, Callable
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Callable,
+)
 
 import paddle
 
 from .... import psdb
 from ....profiler import EventGuard
 from ....utils import (
+    ENV_SOT_ALLOW_DYNAMIC_SHAPE,
     ENV_SOT_EXPORT,
     get_static_function,
     is_break_graph_api,
@@ -33,8 +39,13 @@
     is_not_supported_paddle_layer,
     is_paddle_api,
     magic_method_builtin_dispatch,
+    map_if,
+)
+from ....utils.exceptions import (
+    BreakGraphError,
+    FallbackError,
+    SotErrorBase,
 )
-from ....utils.exceptions import BreakGraphError, FallbackError, SotErrorBase
 from ..dispatcher import Dispatcher
 from ..guard import (
     StringifiedExpression,
@@ -626,11 +637,58 @@ def __init__(
         self.value = fn
 
     def call_function(self, /, *args, **kwargs):
+        from .basic import SymbolicVariable
+
         # Lookup the handler from dispatcher
         handler = Dispatcher.dispatch(self.value, *args, **kwargs)
+
         if handler is not None:
             return handler(*args, **kwargs)
 
+        if ENV_SOT_ALLOW_DYNAMIC_SHAPE.get() and any(
+            isinstance(var, SymbolicVariable)
+            for var in itertools.chain(args, kwargs.values())
+        ):
+            fake_args, fake_kwargs = map_if(
+                (args, kwargs),
+                pred=lambda x: isinstance(x, SymbolicVariable),
+                # this is a fake args, we don't need to care about the value of the args
+                true_fn=lambda x: ConstantVariable.wrap_literal(
+                    None, graph=self.graph
+                ),
+                false_fn=lambda x: x,
+            )
+            handler = Dispatcher.dispatch(self.value, *fake_args, **fake_kwargs)
+            if handler is not None:
+                from ..executor_cache import (
+                    OpcodeExecutorCache,
+                )
+
+                symbolic_inputs = OpcodeExecutorCache().get_symbolic_inputs(
+                    self.graph.pycode_gen._origin_code
+                )
+
+                for var in itertools.chain(args, kwargs.values()):
+                    if isinstance(var, SymbolicVariable):
+                        if var.tracker.is_traceable():
+                            tracker_expr = (
+                                var.tracker.trace_value_from_frame().inlined_expr
+                            )
+                            symbolic_inputs[tracker_expr] = None
+                        else:
+                            for traceable_var in var.get_traceable_inputs():
+                                tracker_expr = (
+                                    traceable_var.tracker.trace_value_from_frame().inlined_expr
+                                )
+                                symbolic_inputs[tracker_expr] = None
+                args, kwargs = map_if(
+                    (args, kwargs),
+                    pred=lambda x: isinstance(x, SymbolicVariable),
+                    true_fn=lambda x: x.to_constant(),
+                    false_fn=lambda x: x,
+                )
+                return handler(*args, **kwargs)
+
         # Try to inline call the magic function
         magic_methods = magic_method_builtin_dispatch(self.value)
         for magic_method in magic_methods:

python/paddle/jit/sot/utils/__init__.py

@@ -32,6 +32,7 @@
 )
 from .exceptions import (  # noqa: F401
     BreakGraphError,
+    DynamicShapeFallbackError,
     ExportError,
     FallbackError,
     InnerError,

python/paddle/jit/sot/utils/exceptions.py

@@ -11,6 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+from __future__ import annotations
 
 import traceback
 
@@ -41,6 +42,10 @@ def __init__(self, msg, disable_eval_frame=False):
         self.disable_eval_frame = disable_eval_frame
 
 
+class DynamicShapeFallbackError(SotErrorBase):
+    pass
+
+
 # raise in inline function call strategy.
 class BreakGraphError(SotErrorBase):
     pass
@@ -52,6 +57,8 @@ def inner_error_default_handler(func, message_fn):
     def impl(*args, **kwargs):
         try:
             return func(*args, **kwargs)
+        except SotErrorBase as e:
+            raise e
         except Exception as e:
             message = message_fn(*args, **kwargs)
             origin_exception_message = "\n".join(

test/sot/test_trace_list_arg.py

@@ -62,16 +62,15 @@ def test_bar_static_shape(self):
 
     @with_allow_dynamic_shape_guard(True)
     def test_bar_dynamic_shape(self):
-        # TODO: Fix this after implement symbolic fallback mechanism
         a = [paddle.to_tensor(1), paddle.to_tensor(2), paddle.to_tensor(3)]
         b = [paddle.to_tensor([2, 3]), paddle.to_tensor(4), paddle.to_tensor(5)]
         with test_instruction_translator_cache_context() as cache:
             self.assert_results(bar, a, 1, 1)
             self.assertEqual(cache.translate_count, 1)
             self.assert_results(bar, a, 2, 0)  # Cache hit, but break graph
-            self.assertEqual(cache.translate_count, 3)
+            self.assertEqual(cache.translate_count, 2)
             self.assert_results(bar, b, 1, 1)  # Cache hit
-            self.assertEqual(cache.translate_count, 3)
+            self.assertEqual(cache.translate_count, 2)
 
 
 if __name__ == "__main__":