Add xla random generator. #9539
Conversation
|
@qihqi please review |
|
a kindly ping @qihqi |
|
Just to understand: so the GeneratorImpl doesn't actually implement the RNG algorithm only manages the state? |
|
|
||
| auto state_tensor = at::empty({(int64_t)total_size}, at::TensorOptions().dtype(at::kByte).device(at::kCPU)); | ||
| uint8_t* data_ptr = state_tensor.data_ptr<uint8_t>(); | ||
| memcpy(data_ptr, &state_->seed_, seed_size); |
There was a problem hiding this comment.
nit: can probably write to state_tensor.index(0).fill_( ...) to write to the tensor.
that is correct. it is exactly what https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/cuda/CUDAGeneratorImpl.h does. the idea is that the generator manages the state and different RNG algorithm generates the random numbers based on the state obtained from the generator. |
|
@qihqi is there anyway to auto-format the cpp code? i know we use |
nvm, i found the linter uses |
|
it seems that the failed tests are not related to this change. i've formatted the files. I don't have the permission to run the tests, please trigger them for me. also, merge if it LGTY. thanks! @qihqi |
This is the very first PR for #9159. It purely add the generator without any utilization of it. #9159 (comment) comment outlines the steps for entire change.