support of Torch > 1.1 about coco-lm HOT 3 CLOSED

Hi,

I'm not familiar with the package compatibility of the functions there. However, I have used pytorch > 1.7.0 with no problem running the fairseq code version (with fused_ops). Could you simply try running with some later pytorch version and see if there are any errors reported?

Best,
Yu

from coco-lm.

Sorry, the use of THC is not compatible with torch>1.11, instead 1.1. The problem I met is the same as HERE.
I managed to compile the fused_ops with torch==1.11, and found that I would succeed if I made the following changes.

1. remove all include about "THC" in *.cu files.
2. change THCudaCheck(cudaGetLastError()) to AT_CUDA_CHECK(cudaGetLastError()).
3. change the include of THC/THCDeviceUtils.cuh to ATen/cuda/DeviceUtils.cuh.
4. remove extern THCState *state in softmax_dropout_kernel.cu.

from coco-lm.

Hi,

Thanks for providing more details. I have limited knowledge about the THC part so I couldn't provide any useful information about it. I'm not sure if making those changes will somehow change the code behavior (hopefully not). The safest way probably is to create a virtual environment and run with pytorch <= 1.9.0 (I remembered this is the latest version that I have tested). Thank you anyways for providing the information for making the package compatible with newer pytorch versions.

Best,
Yu

from coco-lm.