Remove torch cuda empty cache#5678
Draft
shurkat-nvidia wants to merge 2 commits into
Draft
Conversation
shurkat-nvidia
force-pushed
the
remove-torch-cuda-empty-cache
branch
from
adbf098 to
ddbde1f
Compare
torch.cuda.empty_cache() was added as a workaround for out-of-memory errors on some models. With async-shm checkpointing, some of the memory pressure is reduced. Remaining memory fragmentation and pressure can be reduced by properly configuring the torch allocator. Signed-off-by: Skand Hurkat <shurkat@nvidia.com>
- Update checkpointing to use nvrx with shm. - Provide PYTORCH_ALLOC_CONF environment variable to initiate garbage collection. Use larger blocks to avoid fragmentation due to splitting. Signed-off-by: Skand Hurkat <shurkat@nvidia.com>
shurkat-nvidia
force-pushed
the
remove-torch-cuda-empty-cache
branch
from
ddbde1f to
da16039
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
The PR removes an expensive
torch.cuda.empty_cache()call withinsave_checkpoint_and_time(). This call took multiple seconds on every checkpoint and also increased the size of thegenerate_state_dict()function by multiple seconds. To avoid OOM errors, the code also modifies the Nemotron3 Super script to 1) use SHM to avoid expensiverebuild_cuda_tensor()calls during the IPC handoff and 2) garbage collect the torch allocator at 80% GPU allocation with a lower overhead than a rawempty_cache()call, and to use larger blocks to prevent fragmentation in the first place.Issue tracking
For PRs from open-source community contributors:
Linked issue:
Contribution process
Pre-checks
Code review
Feel free to message or comment @NVIDIA/mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!
All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.
Step 1: Mark PR as "Ready for Review"
.github/CODEOWNERS.Final Review might get declined if these requirements are not fulfilled.
Step 2: Final Review
For PRs that change
megatron/core, once all expert reviewers have approved, theFinal Reviewlabel is applied automatically and final reviewers are assigned.For PRs outside
megatron/core, this step is skipped.Step 3: Approved
Once all required reviewers have approved, the
Approvedlabel is applied automatically.Merge
Any member of mcore-engineers will be able to merge your PR.