🐛 [Bug] nn.BatchNorm3d is not compiling (+ can't disable fast_partitioner)

[orioninthesky98]

Bug Description

I can't compile this model and the error seems to be caused by nn.BatchNorm3d

To Reproduce

Steps to reproduce the behavior:

1. Init the model after importing the classes in this gist: https://gist.github.com/orioninthesky98/8a2012f555b7bd4ce50398ff2a1c9291

conv_block = ConvBlock(
    in_channels=1,
    out_channels=16,
    kernel_size=[1, 1, 3],
    pool_ksize=[1, 1, 2],
)
conv_block = conv_block.to("cuda")
conv_block.eval()

3. Compile it with this command

batch_size = 128
network_input_shape = [1, 1, 1, 32]

placeholder_batch = torch.rand((batch_size,) + tuple(network_input_shape))
placeholder_batch = placeholder_batch.to("cuda")

compiled_model = trt.compile(
    conv_block,
    inputs=[placeholder_batch],
    enabled_precisions={torch.float32},
    optimization_level=5,  # max is 5, compilation takes longer but gives the best speedup
    debug=True,  # very verbose, only turn on if needed
    use_fast_partitioner=True,  #  cant disable, results in error when exporting
    dynamic=False,
    disable_tf32=True,  # reduce precision errors at the expense of small slowdown
)

Get the below error (here for full trace: https://gist.github.com/orioninthesky98/9e51a9e83232aa3cac64ce68fe0e512b)

DEBUG:torch_tensorrt.dynamo.conversion._TRTInterpreter:Converting node /cc2_norm/_native_batch_norm_legit_no_training_1 (kind: aten._native_batch_norm_legit_no_training.default, args: ('[CONVOLUTION]-[aten_ops.convolution.default]-[/cc2_conv/convolution_1]_output <tensorrt.ITensor [shape=(81), dtype=DataType.FLOAT]>', '<torch.Tensor as np.ndarray [shape=(16,), dtype=float32]>', '<torch.Tensor as np.ndarray [shape=(16,), dtype=float32]>', '<torch.Tensor as np.ndarray [shape=(16,), dtype=float32]>', '<torch.Tensor as np.ndarray [shape=(16,), dtype=float32]>', 0.1, 1e-05))
ERROR:torch_tensorrt [TensorRT Conversion Context]:ITensor::getDimensions: Error Code 4: Internal Error (Output shape can not be computed for node [CONVOLUTION]-[aten_ops.convolution.default]-[/cc2_conv/convolution_1].)

File ~/.conda/lib/python3.10/site-packages/torch_tensorrt/dynamo/conversion/aten_ops_converters.py:111, in aten_ops_batch_norm_legit_no_training(ctx, target, args, kwargs, name)
     99 @dynamo_tensorrt_converter(
    100     torch.ops.aten._native_batch_norm_legit_no_training.default,
    101     capability_validator=one_user_validator,
   (...)
    109     name: str,
    110 ) -> Union[TRTTensor, Sequence[TRTTensor]]:
--> 111     return impl.normalization.batch_norm(
    112         ctx,
    113         target,
    114         SourceIR.ATEN,
    115         name,
    116         input=args[0],
    117         weight=args[1],
    118         bias=args[2],
    119         running_mean=args[3],
    120         running_var=args[4],
    121         training=False,
    122         momentum=args[5],
    123         eps=args[6],
    124         cudnn_enabled=False,
    125         return_mean_rstd=(
    126             target == torch.ops.aten._native_batch_norm_legit_no_training.default
    127         ),
    128     )

File ~/.conda/lib/python3.10/site-packages/torch_tensorrt/dynamo/conversion/impl/normalization/ops.py:65, in batch_norm(ctx, target, source_ir, name, input, weight, bias, running_mean, running_var, training, momentum, eps, cudnn_enabled, return_mean_rstd)
     63 # For BatchNorm1d, reshape 1d to 2d
     64 output_shape = input.shape
---> 65 if len(input.shape) < 4:
     66     assert (
     67         len(get_dynamic_dims(input.shape)) <= 1
     68     ), "BatchNorm1D with more than one dynamic dims is not currently supported."
     69     new_shape = (
     70         (input.shape[0], input.shape[1], 1, 1)
     71         if len(input.shape) == 2
     72         else (input.shape[0], input.shape[1], input.shape[2], 1)
     73     )

ValueError: __len__() should return >= 0

While executing %_native_batch_norm_legit_no_training_1 : [num_users=1] = call_function[target=torch.ops.aten._native_batch_norm_legit_no_training.default](args = (%convolution_1, %cc2_norm_weight, %cc2_norm_bias, %cc2_norm_running_mean, %cc2_norm_running_var, 0.1, 1e-05), kwargs = {_itensor_to_tensor_meta: {<tensorrt_bindings.tensorrt.ITensor object at 0x7f982e80c6b0>: ((128, 1, 1, 1, 32), torch.float32, False, (32, 32, 32, 32, 1), torch.contiguous_format, False, {}), <tensorrt_bindings.tensorrt.ITensor object at 0x7f982e8175f0>: ((128, 1, 1, 1, 32), torch.float32, False, (32, 32, 32, 32, 1), torch.contiguous_format, False, {}), <tensorrt_bindings.tensorrt.ITensor object at 0x7f982e528eb0>: ((128, 16, 1, 1, 32), torch.float32, False, (512, 32, 32, 32, 1), torch.contiguous_format, False, {}), <tensorrt_bindings.tensorrt.ITensor object at 0x7f982e938c70>: ((128, 16, 1, 1, 32), torch.float32, False, (512, 32, 32, 32, 1), torch.contiguous_format, False, {}), <tensorrt_bindings.tensorrt.ITensor object at 0x7f982e7ce770>: ((128, 16, 1, 1, 32), torch.float32, False, (512, 32, 32, 32, 1), torch.contiguous_format, False, {}), <tensorrt_bindings.tensorrt.ITensor object at 0x7f982e96ef30>: ((128, 16, 1, 1, 32), torch.float32, False, (512, 32, 32, 32, 1), torch.contiguous_format, False, {})}})
Original traceback:
    from loguru import logger
    return forward_call(*args, **kwargs)
    h = self.norm(h)

Expected behavior

The model should compile successfully.

Environment

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

• Torch-TensorRT Version (e.g. 1.0.0): 2.4.0
• PyTorch Version (e.g. 1.0): '2.4.1+cu121'
• CPU Architecture: x86_64
• OS (e.g., Linux): Linux, "Ubuntu 22.04.4 LTS"
• How you installed PyTorch (conda, pip, libtorch, source): conda + pip
• Build command you used (if compiling from source): N/A
• Are you using local sources or building from archives: N/A
• Python version: 3.10
• CUDA version: 12.5
• GPU models and configuration: 1 x H100
• Any other relevant information: N/A

Additional context

update:
I have found a temporary hotfix by replacing BatchNorm3d with BatchNorm1d. This works for some of our use-cases but we still have many models which do need the full 3d BatchNorm, so a fix would still be greatly appreciated.

1. replace this line with self.norm = ( nn.BatchNorm1d(in_channels) if self.is_full_preact else nn.BatchNorm1d(out_channels) )
   https://gist.github.com/orioninthesky98/8a2012f555b7bd4ce50398ff2a1c9291#file-model-py-L74
2. replace this line https://gist.github.com/orioninthesky98/8a2012f555b7bd4ce50398ff2a1c9291#file-model-py-L108-L109 with:

                if self.norm_type == "sn":
                    pass
                elif self.norm_type == "bn":
                    # hotfix for tensorrt compile: squeeze extra 2 dims: (N,C,D,H,W) -> (N,C,H,W) -> (N,C,W)
                    h = h.squeeze(2).squeeze(2)
                    h = self.norm(h)
                    h = h.unsqueeze(2).unsqueeze(2)
                else:
                    h = self.norm(h)

log: https://gist.github.com/orioninthesky98/96612bfd59e35344182de44d9a303aa7

related bug:
if I try to set use_fast_partitioner=False, the model actually compiles fine, but I get this error at the very end and the script crashes, https://gist.github.com/orioninthesky98/a784c361ebbdfa9000564b3f8a1ac1c0) somebody already filed this bug: #3157

[orioninthesky98]

any updates on this?

[lanluo-nvidia]

@orioninthesky98
I have tried the example in the current latest main and our upcoming 2.5.0 release, both are working as expected.
I think the batchnorm3d bug has been fixed.

Also in terms of the use_fast_partitioner=False, bug: #3157 PR is raised and will be merged into main and 2.5.0 release.