Adds a heuristic upper bound in the case of unbounded symints

[narendasan]

Description

Repeated calls to torch.compile without explicit marking of dynamic dimensions may produce unbounded symints (no max val). We need a heuristic and throw a warning to unblock users.

We now assume the upper bound is 2**16 for any particular dim (so that overflow would be less likely for tensor volume calculations done in int32).

Added a bunch of test cases and integrated it into partitioning to generate intermediate shape ranges.

Supersedes: #4080 cc: @wenbingl

Type of change

Please delete options that are not relevant and/or add your own.

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/pass_manager.py	2026-02-17 23:34:29.103514+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/pass_manager.py	2026-02-17 23:35:05.140759+00:00
@@ -123,13 +123,13 @@
        self._validated = False

    def check_pass_names_valid(self, debug_pass_names: List[str]) -> None:
        pass_names_str = [p.__name__ for p in self.passes]
        for name in debug_pass_names:
-            assert name in pass_names_str, (
-                f"{name} is not a valid pass! Passes: {pass_names_str}"
-            )
+            assert (
+                name in pass_names_str
+            ), f"{name} is not a valid pass! Passes: {pass_names_str}"

    def __call__(self, gm: Any, settings: CompilationSettings) -> Any:
        self.validate()
        out = gm
        for _pass in self.passes:
--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/models/test_unbounded_symint.py	2026-02-17 23:34:29.129089+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/models/test_unbounded_symint.py	2026-02-17 23:35:11.671782+00:00
@@ -27,20 +27,19 @@
    class SimpleModel(torch.nn.Module):
        def __init__(self):
            super().__init__()
            self.linear1 = torch.nn.Linear(128, 64)

-
        def forward(self, x):
            return F.relu(self.linear1(x))

    model = SimpleModel().eval().cuda()

    # Create input with unbounded batch dimension
    input_tensor = torch.randn(4, 128).cuda()
    # Mark dimension 0 as dynamic with only min (no max = unbounded)
-    #torch._dynamo.mark_dynamic(input_tensor, 0, min=1)
+    # torch._dynamo.mark_dynamic(input_tensor, 0, min=1)

    compile_spec = {
        "device": torchtrt.Device("cuda:0"),
        "enabled_precisions": {torch.float},
        "min_block_size": 1,
@@ -246,17 +245,15 @@

    input_tensor = torch.randn(8, 16, 16).cuda()
    output_ref = model(input_tensor)
    output_trt = trt_model(input_tensor)

-
    cos_sim = cosine_similarity(output_ref, output_trt)
    assertions.assertTrue(
        cos_sim > COSINE_THRESHOLD,
        msg=f"Reshape with unbounded SymInt test failed. Cosine sim: {cos_sim}",
    )
-

    # Verify output shapes match
    assertions.assertEqual(output_ref.shape, output_trt.shape)

    torch._dynamo.reset()
@@ -360,10 +357,11 @@
            msg=f"Reasonable default test failed at batch_size={batch_size}. Cosine sim: {cos_sim}",
        )

    torch._dynamo.reset()

+
@pytest.mark.unit
def test_unbounded_symint_fallback():
    """
    Test that the default max (min * 128) is applied for unbounded SymInts.
    This test verifies the fallback behavior when no max is specified.
@@ -374,11 +372,10 @@
            super().__init__()
            self.linear1 = torch.nn.Linear(64, 128)
            self.linear2 = torch.nn.Linear(128, 64)
            self.linear3 = torch.nn.Linear(64, 32)
            self.relu = torch.nn.ReLU()
-

        def forward(self, x):
            x = self.relu(self.linear1(x))
            x = self.relu(self.linear2(x))
            x = self.linear3(x)
@@ -412,8 +409,7 @@
        )

    torch._dynamo.reset()


-
if __name__ == "__main__":
    pytest.main([__file__, "-v"])

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/pass_manager.py	2026-02-17 23:34:29.128708+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/pass_manager.py	2026-02-17 23:35:05.630531+00:00
@@ -123,13 +123,13 @@
        self._validated = False

    def check_pass_names_valid(self, debug_pass_names: List[str]) -> None:
        pass_names_str = [p.__name__ for p in self.passes]
        for name in debug_pass_names:
-            assert name in pass_names_str, (
-                f"{name} is not a valid pass! Passes: {pass_names_str}"
-            )
+            assert (
+                name in pass_names_str
+            ), f"{name} is not a valid pass! Passes: {pass_names_str}"

    def __call__(self, gm: Any, settings: CompilationSettings) -> Any:
        self.validate()
        out = gm
        for _pass in self.passes:
--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/models/test_unbounded_symint.py	2026-02-17 23:34:29.154709+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/models/test_unbounded_symint.py	2026-02-17 23:35:11.995372+00:00
@@ -27,20 +27,19 @@
    class SimpleModel(torch.nn.Module):
        def __init__(self):
            super().__init__()
            self.linear1 = torch.nn.Linear(128, 64)

-
        def forward(self, x):
            return F.relu(self.linear1(x))

    model = SimpleModel().eval().cuda()

    # Create input with unbounded batch dimension
    input_tensor = torch.randn(4, 128).cuda()
    # Mark dimension 0 as dynamic with only min (no max = unbounded)
-    #torch._dynamo.mark_dynamic(input_tensor, 0, min=1)
+    # torch._dynamo.mark_dynamic(input_tensor, 0, min=1)

    compile_spec = {
        "device": torchtrt.Device("cuda:0"),
        "enabled_precisions": {torch.float},
        "min_block_size": 1,
@@ -246,17 +245,15 @@

    input_tensor = torch.randn(8, 16, 16).cuda()
    output_ref = model(input_tensor)
    output_trt = trt_model(input_tensor)

-
    cos_sim = cosine_similarity(output_ref, output_trt)
    assertions.assertTrue(
        cos_sim > COSINE_THRESHOLD,
        msg=f"Reshape with unbounded SymInt test failed. Cosine sim: {cos_sim}",
    )
-

    # Verify output shapes match
    assertions.assertEqual(output_ref.shape, output_trt.shape)

    torch._dynamo.reset()
@@ -360,10 +357,11 @@
            msg=f"Reasonable default test failed at batch_size={batch_size}. Cosine sim: {cos_sim}",
        )

    torch._dynamo.reset()

+
@pytest.mark.unit
def test_unbounded_symint_fallback():
    """
    Test that the default max (min * 128) is applied for unbounded SymInts.
    This test verifies the fallback behavior when no max is specified.
@@ -374,11 +372,10 @@
            super().__init__()
            self.linear1 = torch.nn.Linear(64, 128)
            self.linear2 = torch.nn.Linear(128, 64)
            self.linear3 = torch.nn.Linear(64, 32)
            self.relu = torch.nn.ReLU()
-

        def forward(self, x):
            x = self.relu(self.linear1(x))
            x = self.relu(self.linear2(x))
            x = self.linear3(x)
@@ -412,8 +409,7 @@
        )

    torch._dynamo.reset()


-
if __name__ == "__main__":
    pytest.main([__file__, "-v"])

old

[zewenli98]

kind of confused about the test cases. Shouldn't we mark them explicitly?

[zewenli98]

isn't it bounded?

[zewenli98]

should mark here?