support native int64 as valid datatype

pytorch / TensorRT

PyTorch/TorchScript/FX compiler for NVIDIA GPUs using TensorRT

https://pytorch.org/TensorRT

BSD 3-Clause "New" or "Revised" License

2.41k stars 333 forks source link

support native int64 as valid datatype #2937

Open apbose opened 2 weeks ago

HolyWu commented 6 days ago

It seems that https://github.com/pytorch/TensorRT/pull/2855/files#diff-97c67f0cea25ec7aa88ca61594d9ad62e9fb575cab3647b079cb718f81308454 is already planned to do relevant changes.

apbose commented 6 days ago

Thanks @HolyWu for pointing out the above. @narendasan would the changes in PR https://github.com/pytorch/TensorRT/pull/2855/files#diff-97c67f0cea25ec7aa88ca61594d9ad62e9fb575cab3647b079cb718f81308454 work or would we have to create a type map for kINT32 to kINT64 in the torchscript path?

narendasan commented 6 days ago

Yeah neither of these PRs implements the feature fully since we need to either amend truncation in TorchScript or have a TorchScript specific type map