PR #7318 introduced OpenXLA CUDA fallback. However, that's an opt-in feature that relies on the environment variable: XLA_FALLBACK_CUDA. Even though that works fine, as this comment suggests, it doesn't provide a good UX. Therefore, we should improve it further by adding the following changes:
[x] Make OpenXLA CUDA fallback the default fallback, making the environment negating the environment variable to something like DISABLE_XLA_FALLBACK_CUDA #7630
[ ] Deprecate the environment variable, by making PyTorch/XLA smart enough to pick the appropriate fallback
PR #7318 introduced OpenXLA CUDA fallback. However, that's an opt-in feature that relies on the environment variable:
XLA_FALLBACK_CUDA
. Even though that works fine, as this comment suggests, it doesn't provide a good UX. Therefore, we should improve it further by adding the following changes:DISABLE_XLA_FALLBACK_CUDA
#7630cc @miladm @JackCaoG @vanbasten23