🐛 [Bug] opt-shape cannot be set same as min-shape when torch_tensorrt.compile

[kevinteng525]

Bug Description

torch_tensorrt.compile will failed if setting opt-shape same as min-shape

To Reproduce

Steps to reproduce the behavior:

1. set opt-shape same as min-shape as below

model_compiled = torch_tensorrt.compile(
    model,
    inputs=[
        torch_tensorrt.Input(
            min_shape=[1, 3, 256, 256],
            opt_shape=[1, 3, 256, 256],
            max_shape=[128, 3, 256, 256],
            dtype=input.dtype
        )
        ],
    )

2. compile failed:
   `
   File "/usr/local/lib/python3.12/site-packages/torch/export/exported_program.py", line 124, in wrapper
   return fn(*args, **kwargs)
   ^^^^^^^^^^^^^^^^^^^
   File "/usr/local/lib/python3.12/site-packages/torch/export/_trace.py", line 2071, in _export_for_training
   export_artifact = export_func(
   ^^^^^^^^^^^^
   File "/usr/local/lib/python3.12/site-packages/torch/export/_trace.py", line 2002, in _non_strict_export
   aten_export_artifact = _to_aten_func( # type: ignore[operator]
   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   File "/usr/local/lib/python3.12/site-packages/torch/export/_trace.py", line 1823, in _export_to_aten_ir_make_fx
   raise UserError(UserErrorType.CONSTRAINT_VIOLATION, str(e)) # noqa: B904
   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   torch._dynamo.exc.UserError: Constraints violated (_0)! For more information, run with TORCH_LOGS="+dynamic".

• You marked _0 as dynamic but your code specialized it to be a constant (1). If you're using mark_dynamic, either remove it or use maybe_mark_dynamic. If you're using Dim.DYNAMIC, replace it with either Dim.STATIC or Dim.AUTO.
  Suggested fixes:
  _0 = 1
  The error above occurred when calling torch.export.export. If you would like to view some more information about this error, and get a list of all other errors that may occur in your export call, you can replace your export() call with draft_export().
  `

Expected behavior

Compile success

Environment

Build information about Torch-TensorRT can be found by turning on debug messages

• Torch-TensorRT Version (e.g. 1.0.0): 2.10.0
• PyTorch Version (e.g. 1.0): 2.9.0
• CPU Architecture: x86_64
• OS (e.g., Linux): Linux
• Python version: 3.12
• CUDA version: 12.9
• GPU models and configuration: H20

[narendasan]

This seems like a PyTorch issue fundamentally. For example, if I have the following program:

import torch
import torch_tensorrt as torchtrt


# Simple model
class MLP(torch.nn.Module):
    def forward(self, x):
        return torch.relu(x @ x.T)


model = MLP().eval().cuda()

# Define inputs with dynamic batch dim
inputs = [torch.randn(1, 128, device="cuda")] # An example input that is the same shape as min shape

dynamic_shapes = ({0: torch.export.Dim("batch", min=1, max=32)},) # Explicitly set the dynamic shape for tracing 

# Export with dynamic shapes
exp_program = torch.export.export(model, tuple(inputs), dynamic_shapes=dynamic_shapes) # This line fails 

You get a similar error as the one you reported. The reason is torch_tensorrt.compile is a simple wrapper around torch.export.export and torch_tensorrt.dynamo.compile, where we use opt shape as the example input for torch export. Simply changing the example input to something else passes this step. I will forward this issue to PyTorch but to unblock you can WAR by separating the implicit steps inside torch_tensorrt.compile. Opt shape is meaningful for TensorRT to set the shape we should prefer to tune for but is less important for the export phase. Therefore you can set some arbitrary non min shape for export and then set opt shape as you want for the torch_tensorrt compilation phase.

e.g.

import torch
import torch_tensorrt as torchtrt


# Simple model
class MLP(torch.nn.Module):
    def forward(self, x):
        return torch.relu(x @ x.T)


model = MLP().eval().cuda()

# Define inputs with dynamic batch dim
inputs = [torch.randn(4, 128, device="cuda")]

dynamic_shapes = ({0: torch.export.Dim("batch", min=1, max=32)},)

# Export with dynamic shapes
exp_program = torch.export.export(model, tuple(inputs), dynamic_shapes=dynamic_shapes)

# Compile with TensorRT, providing min/opt/max input specs
trt_inputs = [
    torchtrt.Input(
        min_shape=(1, 128),
        opt_shape=(1, 128),
        max_shape=(32, 128),
        dtype=torch.float32,
    )
]

trt_model = torchtrt.dynamo.compile(exp_program, inputs=trt_inputs)

# Run inference at different batch sizes
for batch in [1, 8, 32]:
    x = torch.randn(batch, 128, device="cuda")
    out = trt_model(x)
    print(f"batch={batch}, output shape: {out.shape}")