WIP Support python slicing with data depedennt inptu tensors maybe

ghstack-source-id: 5363cf5565c2c024260b6ae504ba12ffca2f9984
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165074

test/dynamo/test_error_messages.py

@@ -528,30 +528,6 @@ Attempted to call function marked as skipped
             f(x)
         self.assertEqual(len(ws), 2)
 
-    def test_slice_with_tensor(self):
-        def fn(x, y):
-            return x[:y]
-
-        self.assertExpectedInlineMunged(
-            Unsupported,
-            lambda: torch.compile(fn, backend="eager", fullgraph=True)(
-                torch.randn(10),
-                torch.tensor([3]),
-            ),
-            """\
-Dynamic slicing with Tensor arguments
-  Explanation: Creating slices with Tensor arguments is not supported. e.g. `l[:x]`, where `x` is a 1-element tensor.
-  Hint: It may be possible to write Dynamo tracing rules for this code. Please report an issue to PyTorch if you encounter this graph break often and it is causing performance issues.
-
-  Developer debug context: SliceVariable start: ConstantVariable(NoneType: None), stop: LazyVariableTracker(realized: TensorVariable()), step: ConstantVariable(NoneType: None)
-
- For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0038.html
-
-from user code:
-   File "test_error_messages.py", line N, in fn
-    return x[:y]""",
-        )
-
     def test_observed_exception(self):
         def fn():
             raise RuntimeError("test")

test/test_dynamic_shapes.py

@@ -3799,6 +3799,118 @@ def forward(self, arg0_1: "i64[2][1]cpu", arg1_1: "Sym(u2)", arg2_1: "Sym(u3)",
     def test_unbacked_slice_with_step_cpp_wrapper(self):
         self.test_unbacked_slice_with_step()
 
+    @fresh_cache()
+    @torch._dynamo.config.patch("capture_scalar_outputs", True)
+    def test_slice_with_tensor_indices(self):
+        # Test slicing with tensor start/stop/step on RHS (reading)
+
+        # Test 1: Basic slice with tensor start and stop
+        def f1(x, start_t, stop_t):
+            return x[start_t:stop_t]
+
+        x = torch.randn(20)
+        start_t = torch.tensor(5)
+        stop_t = torch.tensor(15)
+        fn1 = torch.compile(f1, fullgraph=True, backend="inductor")
+        self.assertTrue(torch.allclose(fn1(x, start_t, stop_t), f1(x, start_t, stop_t)))
+
+        # Test 2: Slice with tensor step
+        def f2(x, start_t, stop_t, step_t):
+            return x[start_t:stop_t:step_t]
+
+        step_t = torch.tensor(2)
+        fn2 = torch.compile(f2, fullgraph=True, backend="inductor")
+        self.assertTrue(
+            torch.allclose(
+                fn2(x, start_t, stop_t, step_t), f2(x, start_t, stop_t, step_t)
+            )
+        )
+
+        # Test 3: Slice with only tensor start
+        def f3(x, start_t):
+            return x[start_t:]
+
+        fn3 = torch.compile(f3, fullgraph=True, backend="inductor")
+        self.assertTrue(torch.allclose(fn3(x, start_t), f3(x, start_t)))
+
+        # Test 4: Slice with only tensor stop
+        def f4(x, stop_t):
+            return x[:stop_t]
+
+        fn4 = torch.compile(f4, fullgraph=True, backend="inductor")
+        self.assertTrue(torch.allclose(fn4(x, stop_t), f4(x, stop_t)))
+
+        # Test 5: Negative indices with tensors
+        def f5(x, start_t):
+            return x[start_t:-1]
+
+        start_t_neg = torch.tensor(-10)
+        fn5 = torch.compile(f5, fullgraph=True, backend="inductor")
+        self.assertTrue(torch.allclose(fn5(x, start_t_neg), f5(x, start_t_neg)))
+
+        # Test 6: Multidimensional slice with tensor indices
+        def f6(x, start_t, stop_t):
+            return x[:, start_t:stop_t]
+
+        x_2d = torch.randn(10, 20)
+        fn6 = torch.compile(f6, fullgraph=True, backend="inductor")
+        self.assertTrue(
+            torch.allclose(fn6(x_2d, start_t, stop_t), f6(x_2d, start_t, stop_t))
+        )
+
+    @fresh_cache()
+    @torch._dynamo.config.patch("capture_scalar_outputs", True)
+    @torch._inductor.config.patch("cpp_wrapper", True)
+    def test_slice_with_tensor_indices_cpp_wrapper(self):
+        self.test_slice_with_tensor_indices()
+
+    @fresh_cache()
+    @torch._dynamo.config.patch("capture_scalar_outputs", True)
+    def test_select_with_tensor_index(self):
+        # Test direct tensor indexing (select) without calling .item()
+
+        # Test 1: Simple 0-d tensor as index
+        def f1(x, idx_tensor):
+            return x[idx_tensor]
+
+        x = torch.randn(10)
+        idx_tensor = torch.tensor(5)
+        fn1 = torch.compile(f1, fullgraph=True, backend="inductor")
+        self.assertTrue(torch.allclose(fn1(x, idx_tensor), f1(x, idx_tensor)))
+
+        # Test 2: Negative tensor index
+        def f2(x, idx_tensor):
+            return x[idx_tensor]
+
+        idx_tensor_neg = torch.tensor(-2)
+        fn2 = torch.compile(f2, fullgraph=True, backend="inductor")
+        self.assertTrue(torch.allclose(fn2(x, idx_tensor_neg), f2(x, idx_tensor_neg)))
+
+        # TODO support those less common patterns
+        # # Test 3: Multidimensional select with tensor index
+        # def f3(x, idx_tensor):
+        #     return x[:, idx_tensor]
+
+        # x_2d = torch.randn(5, 10)
+        # fn3 = torch.compile(f3, fullgraph=True, backend="inductor")
+        # self.assertTrue(torch.allclose(fn3(x_2d, idx_tensor), f3(x_2d, idx_tensor)))
+
+        # Test 4: Multiple tensor indices
+        # def f4(x, idx1, idx2):
+        #     return x[idx1, idx2]
+
+        # x_2d = torch.randn(8, 12)
+        # idx1 = torch.tensor(3)
+        # idx2 = torch.tensor(7)
+        # fn4 = torch.compile(f4, fullgraph=True, backend="inductor")
+        # self.assertTrue(torch.allclose(fn4(x_2d, idx1, idx2), f4(x_2d, idx1, idx2)))
+
+    @fresh_cache()
+    @torch._dynamo.config.patch("capture_scalar_outputs", True)
+    @torch._inductor.config.patch("cpp_wrapper", True)
+    def test_select_with_tensor_index_cpp_wrapper(self):
+        self.test_select_with_tensor_index()
+
     @fresh_cache()
     @torch._dynamo.config.patch("capture_scalar_outputs", True)
     def test_tensor_split(self):

torch/_dynamo/symbolic_convert.py

@@ -3138,7 +3138,7 @@ class InstructionTranslatorBase(
 
     def BUILD_SLICE(self, inst: Instruction) -> None:
         items = self.popn(inst.argval)
-        self.push(SliceVariable(items))
+        self.push(SliceVariable(items, tx=self))
 
     def BUILD_LIST(self, inst: Instruction) -> None:
         items = self.popn(inst.argval)

torch/_dynamo/variables/builder.py

@@ -1795,7 +1795,7 @@ class VariableBuilder:
         ]
         self.install_guards(GuardBuilder.TYPE_MATCH)
         if isinstance(value, slice):
-            return SliceVariable(items, source=self.source)
+            return SliceVariable(items, self.tx, source=self.source)
         else:
             return RangeVariable(items, source=self.source)
 

torch/_dynamo/variables/builtin.py

@@ -1235,6 +1235,20 @@ class BuiltinVariable(VariableTracker):
                 # scenario is to de-sugar eagerly.
                 fn, args = IN_PLACE_DESUGARING_MAP[fn], [args[0], args[1]]
 
+            # Convert size-1 TensorVariable indices to SymIntVariable by calling .item()
+            # This decomposes tensor[t] to u=t.item(); tensor[u] at the dynamo level
+            if (
+                fn is operator.getitem
+                and len(args) == 2
+                and isinstance(args[1], variables.TensorVariable)
+            ):
+                tensor_idx = args[1]
+                # Only convert if we know it's size-1 (not for advanced indexing)
+                if tensor_idx.size is not None and all(s == 1 for s in tensor_idx.size):
+                    args = list(args)
+                    args[1] = tensor_idx.call_method(tx, "item", [], {})
+                    args = tuple(args)
+
             if fn is operator.getitem and isinstance(args[1], SymNodeVariable):
                 # Standard indexing will force specialization due to
                 # __index__.  Rewrite as a regular torch op which will
@@ -1745,7 +1759,7 @@ class BuiltinVariable(VariableTracker):
         )
 
     def call_slice(self, tx: "InstructionTranslator", *args):
-        return variables.SliceVariable(args)
+        return variables.SliceVariable(args, tx)
 
     def _dyn_proxy(self, tx: "InstructionTranslator", *args, **kwargs):
         from .builder import wrap_fx_proxy

torch/_dynamo/variables/lists.py

@@ -1386,7 +1386,7 @@ class NamedTupleVariable(TupleVariable):
 
 
 class SliceVariable(VariableTracker):
-    def __init__(self, items, **kwargs) -> None:
+    def __init__(self, items, tx=None, **kwargs) -> None:
         items_to_map = items
         start, stop, step = [variables.ConstantVariable.create(None)] * 3
 
@@ -1399,18 +1399,27 @@ class SliceVariable(VariableTracker):
         else:
             raise AssertionError
 
-        if isinstance(start, variables.TensorVariable) or isinstance(
+        # Convert TensorVariable to SymIntVariable by calling .item()
-            stop, variables.TensorVariable
+        # This decomposes a[:t] to u=t.item(); a[:u] at the dynamo level
-        ):
+        if isinstance(start, variables.TensorVariable):
-            unimplemented_v2(
+            assert tx is not None, (
-                gb_type="Dynamic slicing with Tensor arguments",
+                "tx is required when slice indices are TensorVariables"
-                context=f"SliceVariable start: {start}, stop: {stop}, step: {step}",
-                explanation="Creating slices with Tensor arguments is not supported. "
-                "e.g. `l[:x]`, where `x` is a 1-element tensor.",
-                hints=[
-                    *graph_break_hints.SUPPORTABLE,
-                ],
             )
+            assert start.size is None or all(s == 1 for s in start.size)
+            start = start.call_method(tx, "item", [], {})
+        if isinstance(stop, variables.TensorVariable):
+            assert tx is not None, (
+                "tx is required when slice indices are TensorVariables"
+            )
+            assert stop.size is None or all(s == 1 for s in stop.size)
+            stop = stop.call_method(tx, "item", [], {})
+        if isinstance(step, variables.TensorVariable):
+            assert tx is not None, (
+                "tx is required when slice indices are TensorVariables"
+            )
+            assert step.size is None or all(s == 1 for s in step.size)
+            step = step.call_method(tx, "item", [], {})
+
         self.items = (start, stop, step)
 
         super().__init__(**kwargs)