ValueError: Invalid input type <class 'bool'> encountered when compiling FLUX.1-dev model with Torch-TensorRT

[yachty66]

❓ Question

When trying to compile the FLUX.1-dev model using Torch-TensorRT following the official example/blog post, I'm encountering a ValueError during the torch_tensorrt.dynamo.compile() step. The error suggests there's an issue with input parsing where it's encountering a boolean value that it doesn't know how to handle.

What you have already tried

I'm following the exact steps from the example provided in the documentation (https://pytorch.org/TensorRT/tutorials/_rendered_examples/dynamo/torch_export_flux_dev.html). I've:

1. Successfully loaded the FLUX.1-dev model
2. Defined the dynamic shapes properly
3. Created dummy inputs with the recommended dimensions
4. Successfully exported the model using _export
5. Attempted to compile with Torch-TensorRT using the same parameters shown in the example

The error occurs specifically at the compilation step:

trt_gm = torch_tensorrt.dynamo.compile(
    ep,
    inputs=dummy_inputs,
    enabled_precisions={torch.float32},
    truncate_double=True,
    min_block_size=1,
    use_fp32_acc=True,
    use_explicit_typing=True,
)

Environment

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

• PyTorch Version (e.g., 1.0): 2.6.0
• CPU Architecture:
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, libtorch, source):
• Build command you used (if compiling from source):
• Are you using local sources or building from archives:
• Python version: 3.11.10
• CUDA version: cuda_12.4.r12.4/compiler.34097967_0
• GPU models and configuration: A100
• Any other relevant information:

Additional context

The error message specifically points to an issue with boolean input types:

ValueError: Invalid input type <class 'bool'> encountered in the dynamo_compile input parsing. Allowed input types: {torch_tensorrt.Input, torch.Tensor, list, tuple, dict}

It looks like the return_dict=False parameter in my dummy inputs is causing the issue since it's a boolean value. The example shows that this should be supported, but the error suggests that booleans aren't handled correctly in the input parsing logic.

Full traceback:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/workspace/flux-dev-tensorrt.ipynb Cell 4 line 1
----> <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=0'>1</a> trt_gm = torch_tensorrt.dynamo.compile(
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=1'>2</a>     ep,
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=2'>3</a>     inputs=dummy_inputs,
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=3'>4</a>     enabled_precisions={torch.float32},
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=4'>5</a>     truncate_double=True,
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=5'>6</a>     min_block_size=1,
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=6'>7</a>     use_fp32_acc=True,
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=7'>8</a>     use_explicit_typing=True,
      <a href='vscode-notebook-cell://ssh-remote%2B216.81.245.143/workspace/flux-dev-tensorrt.ipynb#W5sdnNjb2RlLXJlbW90ZQ%3D%3D?line=8'>9</a> )

File /usr/local/lib/python3.11/dist-packages/torch_tensorrt/dynamo/_compiler.py:606, in compile(exported_program, inputs, arg_inputs, kwarg_inputs, device, disable_tf32, assume_dynamic_shape_support, sparse_weights, enabled_precisions, engine_capability, debug, num_avg_timing_iters, workspace_size, dla_sram_size, dla_local_dram_size, dla_global_dram_size, truncate_double, require_full_compilation, min_block_size, torch_executed_ops, torch_executed_modules, pass_through_build_failures, max_aux_streams, version_compatible, optimization_level, use_python_runtime, use_fast_partitioner, enable_experimental_decompositions, dryrun, hardware_compatible, timing_cache_path, lazy_engine_init, cache_built_engines, reuse_cached_engines, engine_cache_dir, engine_cache_size, custom_engine_cache, use_explicit_typing, use_fp32_acc, refit_identical_engine_weights, strip_engine_weights, immutable_weights, enable_weight_streaming, **kwargs)
    603     arg_inputs = [arg_inputs]  # type: ignore
    605 # Prepare torch_trt inputs
--> 606 trt_arg_inputs: Sequence[Input] = prepare_inputs(arg_inputs)
    607 trt_kwarg_inputs: Optional[dict[Any, Any]] = prepare_inputs(kwarg_inputs)
    608 device = to_torch_tensorrt_device(device)

File /usr/local/lib/python3.11/dist-packages/torch_tensorrt/dynamo/utils.py:257, in prepare_inputs(inputs, disable_memory_format_check)
    255 torchtrt_input_list = []
    256 for input_obj in inputs:
--> 257     torchtrt_input = prepare_inputs(
    258         input_obj, disable_memory_format_check=disable_memory_format_check
    259     )
    260     torchtrt_input_list.append(torchtrt_input)
    262 return (
    263     torchtr

[narendasan]

@peri044 @cehongwang Can you please take a look?

[yachty66]

@peri044 @cehongwang Can you please take a look?

Appreciate it, thank you!

[peri044]

@yachty66 Did you try with the latest main branch or the release 2.6 ?

[yachty66]

@yachty66 Did you try with the latest main branch or the release 2.6 ?

I did pip install torch-tensorrt. Should I try pip install --pre torch-tensorrt --index-url https://download.pytorch.org/whl/nightly/cu124?

[yachty66]

@peri044 it's not possible to install the nightly version:

When running the install command for the nightly version pip install --pre torch-tensorrt --index-url https://download.pytorch.org/whl/nightly/cu124, I am getting the following error:

ERROR: Cannot install torch-tensorrt==2.7.0.dev20250119+cu124, torch-tensorrt==2.7.0.dev20250120+cu124, ... because these package versions have conflicting dependencies.

The conflict is caused by:
    torch-tensorrt 2.7.0.dev20250228+cu124 depends on tensorrt<10.8.0 and >=10.7.0.post1
    ...
    torch-tensorrt 2.7.0.dev20250130+cu124 depends on tensorrt-cu12<10.8.0 and >=10.7.0
    ...

[narendasan]

@zewenli98 Can you take a look at this dependency issue?

[cehongwang]

This has already been solved in the latest main branch. You can upgrade to the latest version of Torch-TRT. Alternatively, you can change the last entry of dummy inputs to: "return_dict": torch.tensor(False) to circumvent this issue for now.

[yachty66]

This has already been solved in the latest main branch. You can upgrade to the latest version of Torch-TRT. Alternatively, you can change the last entry of dummy inputs to: "return_dict": torch.tensor(False) to circumvent this issue for now.

When trying "return_dict": torch.tensor(False) I run in the following error:

W0321 23:52:56.241000 1367 torch/fx/experimental/symbolic_shapes.py:6307] failed during evaluate_expr(Eq(u0, 1), hint=None, size_oblivious=False, forcing_spec=False
E0321 23:52:56.242000 1367 torch/fx/experimental/recording.py:299] failed while running evaluate_expr(*(Eq(u0, 1), None), **{'fx_node': False})
---------------------------------------------------------------------------
GuardOnDataDependentSymNode               Traceback (most recent call last)
[<ipython-input-8-232d7c3393f5>](https://yrohqbkz7ib-496ff2e9c6d22116-0-colab.googleusercontent.com/outputframe.html?vrz=colab_20250320-060136_RC00_738743873#) in <cell line: 0>()
     35 }
     36 # This will create an exported program which is going to be compiled with Torch-TensorRT
---> 37 ep = _export(
     38     backbone,
     39     args=(),

26 frames
[/usr/local/lib/python3.11/dist-packages/torch/fx/experimental/symbolic_shapes.py](https://yrohqbkz7ib-496ff2e9c6d22116-0-colab.googleusercontent.com/outputframe.html?vrz=colab_20250320-060136_RC00_738743873#) in _evaluate_expr(self, orig_expr, hint, fx_node, size_oblivious, forcing_spec)
   6491 
   6492                     if not ok:
-> 6493                         raise self._make_data_dependent_error(
   6494                             expr.xreplace(self.var_to_val),
   6495                             expr,

GuardOnDataDependentSymNode: Could not guard on data-dependent expression Eq(u0, 1) (unhinted: Eq(u0, 1)).  (Size-like symbols: none)

Caused by: (_subclasses/functional_tensor.py:294 in __bool__)
For more information, run with TORCH_LOGS="dynamic"
For extended logs when we create symbols, also add TORCHDYNAMO_EXTENDED_DEBUG_CREATE_SYMBOL="u0"
If you suspect the guard was triggered from C++, add TORCHDYNAMO_EXTENDED_DEBUG_CPP=1
For more debugging help, see https://docs.google.com/document/d/1HSuTTVvYH1pTew89Rtpeu84Ht3nQEFTYhAX3Ypa_xJs/edit?usp=sharing

For C++ stack trace, run with TORCHDYNAMO_EXTENDED_DEBUG_CPP=1