chore: Upgrade TensorRT version to TRT 10 EA (#2699) by peri044 · Pull Request #2774 · pytorch/TensorRT (original) (raw)

@peri044

I followed the code changes to address the error caused by broadcasting, and it seems to be related to this issue.

If we perform element-wise operations on two inputs with shapes lhs (2, 2, 14, 1, 5) and rhs (2, 1, 3, 5), the output of the element-wise operation layer should be (2, 2, 14, 3, 5).
However, with the current code, the output of the element-wise operation layer follows the larger rank of lhs_val and becomes (2, 2, 14, 1, 5).

As a result, an error that did not occur on the local machine was found during CI/CD (#2726)

Is it necessary to make adjustments to fix this code?