Repository navigation
torch.compile for training #4
Description
Is your feature request related to a problem? Please describe.
We should leverage torch.compile for training and fprop.
Describe the solution you'd like
Ideally we should be able to torch compile all calls to hf model (train, get_logprobs). The compile cost should be paid once at the init.
FYI there is a known issue with torch.compile + FSDP + BF16 and torch.compile was failing (with torch 2.5.1).
Describe alternatives you've considered
thunder compiler: https://github.com/Lightning-AI/lightning-thunder/ (we should investigate perf benefits of thunder over native compile)
Activity
feat: add GLM-4.7-Flash GRPO release test (NVIDIA-NeMo#4)
feat: add GLM-4.7-Flash GRPO release test (#4)
fix(data-plane): preserve sample order and FLOPs semantics on @dp_dis…
fix(data-plane): preserve sample order and FLOPs semantics on @dp_dis…
4 remaining items
fix(data-plane): preserve sample order and FLOPs semantics on @dp_dis…
refactor(grpo_sync): drop dead batch_cache; make TQPolicy attrs public
chore(grpo): drop stale mypy comments; rename TQPolicy ctor->actor
fix(data-plane): preserve sample order and FLOPs semantics on @dp_dis…
refactor(grpo_sync): drop dead batch_cache; make TQPolicy attrs public
chore(grpo): drop stale mypy comments; rename TQPolicy ctor->actor
fix(data-plane): preserve sample order and FLOPs semantics on @dp_dis…
refactor(grpo_sync): drop dead batch_cache; make TQPolicy attrs public
chore(grpo): drop stale mypy comments; rename TQPolicy ctor->actor
fix(data-plane): preserve sample order and FLOPs semantics on @dp_dis…
refactor(grpo_sync): drop dead batch_cache; make TQPolicy attrs public
chore(grpo): drop stale mypy comments; rename TQPolicy ctor->actor
feat(ft-experiment): finalize 4-setup fault-tolerance experiment configs
docs: enumerate all per-attempt W&B runs for controls in ft_experimen…
fix: address review feedback on inner TP-worker nsys profiling