TArray used for broadcast was limited to be within range [0, 8] on onnxruntime 1.16.3

[CapJunkrat]

Describe the issue

TArray used for broadcast was limited to be within range [0, 8] on onnxruntime 1.16.3, and I wonder if the tensor to be broadcasted can be higher than 8 dim, or if this limit is removed in higher onnxruntime version.

To reproduce

code to generate onnx model:

import torch
import torch.onnx

y = torch.randn(1,1,3,1,1,5,5)

class MinModel(torch.nn.Module):
    def forward(self, x):
        return torch.min(x, y)  # Perform min operation along the last dimension

# Initialize the model
model = MinModel()

# Create dummy input data with the specified shape (1, 3, 3, 3, 3, 5)
dummy_input = torch.randn(1,1,3,1,1,1,1,1,1,1,5,5)

# Export the model to ONNX format
onnx_path = "min_model.onnx"
torch.onnx.export(
    model,
    dummy_input,
    onnx_path,
    input_names=["inputs"],
    output_names=["outputs"],
    opset_version=12
)

print(f"Model has been exported to {onnx_path}")

Inference this onnx model will get this error:

[E:onnxruntime:, sequential_executor.cc:514 ExecuteKernel] Non-zero status code returned while running Min node. Name:'Min_2' Status Message: /onnxruntime_src/onnxruntime/core/providers/cuda/shared_inc/cuda_utils.h:84 void onnxruntime::cuda::TArray<T, capacity>::SetSize(int32_t) [with T = long int; int capacity = 8; int32_t = int] 0 <= size && size <= capacity was false. TArray size must be within range [0, 8]. Actual: 10

Urgency

No response

Platform

Linux

OS Version

centos 7

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.16.3

ONNX Runtime API

Python

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

cuda 11.8

[xadupre]

It is a converting issue. The exporter does not seem to support more than 8 dimensions for a tensor. Do you need 10?

[CapJunkrat]

The exporter does not seem to support more than 8 dimensions for a tensor

Actually I did not receive any error from the export process. I got that error when I tried to inference the exported onnx model with onnxruntime-gpu. Seems like there is a TArray size range limit no larger than 8.

[gramalingam]

It is a converting issue. The exporter does not seem to support more than 8 dimensions for a tensor. Do you need 10?

The error message is from onnxruntime ... why do you say it is a converter issue?