Support ddp_fork strategy with native AMP by attempting NVML-based CUDA availability assessment#14984

Merged

awaelchli merged 15 commits intoLightning-AI:masterfrom speediedan:enable_ddp_fork_strategy_with_native_amp

Oct 5, 2022

+65-6

Commits on Oct 5, 2022

support DDP_FORK strategy with native AMP by attempting NVML-based CUDA availability assessment
speediedan
committed
update CHANGELOG
speediedan
committed
narrow scope of RunIf NVML-based CUDA availability check to a new RunIf condition
speediedan
committed
clarify new runif condition, add skip_windows condition to new test
speediedan
committed
add context manager that safely patches :func:torch.cuda.is_available with its NVML-based version if possible
speediedan
committed
update min_cuda_gpus RunIf condition for both PL and Lite, narrow scope of new ddp_fork test
speediedan
committed
restructure to avoid unnecessary None check

speediedan
and
carmocca
committed
remove unnecessary None assignment

speediedan
and
carmocca
committed
cleanup verbose documentation, revert AMPTestModel location, refine new _patch_cuda_is_available context manager
speediedan
committed
add missing yield
speediedan
committed
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
nit
carmocca
committed
Merge branch 'master' into enable_ddp_fork_strategy_with_native_amp
Borda
authored
since we're not using my original ddp_fork full strategy test anymore, we only need 1 GPU
speediedan
committed
looks like an unrequired import got added to native_amp sometime in the last few hours
speediedan
committed