[graph_trainer] Force cudagraph for MinimalAsyncEP on H100 + Run 4 results#3610
Draft
SherlockNoMad wants to merge 1 commit into
Draft
[graph_trainer] Force cudagraph for MinimalAsyncEP on H100 + Run 4 results#3610SherlockNoMad wants to merge 1 commit into
SherlockNoMad wants to merge 1 commit into
Conversation
This was referenced Jun 10, 2026
meta-cla
Bot
added
the
CLA Signed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stack from ghstack (oldest at bottom):
Enable cudagraph in the dsv3 run script and force cudagraph_pass past the
_grouped_mm '< sm_100' gate in is_cudagraphable, so MinimalAsyncEP captures a
full cudagraph on H100 (sm_90) instead of skipping (312 _grouped_mm nodes).
Records Run 4 in the results doc: forced cudagraph runs without crashing and
gives a stable ~16.5% MFU / ~9k tps (vs ~12.7% with cudagraph skipped).
WARNING: the cudagraph.py change is an unconditional DEBUG override (TODO:
revert / gate behind a flag or relax only the _grouped_mm gate). Forcing past
this safety check is numerically UNVERIFIED -- _grouped_mm may do hidden
CPU<->CUDA copies that replay stale under cudagraph. Needs an eager loss_compare
before trusting.