Skip to content

Bump torch version, make tests deterministic#80

Merged
romeokienzler merged 8 commits into
mainfrom
bump_torch
Jul 1, 2026
Merged

Bump torch version, make tests deterministic#80
romeokienzler merged 8 commits into
mainfrom
bump_torch

Conversation

@romeokienzler

@romeokienzler romeokienzler commented Jun 26, 2026

Copy link
Copy Markdown
Collaborator
  • edit versions
  • fix test
  • adjust CIs for test

Signed-off-by: Romeo Kienzler <[email protected]>
@romeokienzler romeokienzler requested a review from albanpuech June 26, 2026 08:18
romeokienzler and others added 6 commits June 30, 2026 15:48
Co-Authored-By: Claude Opus 4.7 <[email protected]>
Signed-off-by: Romeo Kienzler <[email protected]>
Defensive on Linux per _warn_mp_context_on_linux.

Co-Authored-By: Claude Opus 4.7 <[email protected]>
Signed-off-by: Romeo Kienzler <[email protected]>
Bounds widened/shifted due to combined effect of torch>=2.10 numerics
and --mp_context spawn re-seeding DataLoader workers per run.

Signed-off-by: Romeo Kienzler <[email protected]>
Adds --deterministic [true|warn] to the train subcommand and threads it
through Lightning's Trainer(deterministic=...). The integration tests
now run with --deterministic warn and CUBLAS_WORKSPACE_CONFIG=:4096:8,
which produced Std=0 across 5 seed=0 calibration runs on case14_ieee.

OPF CI bounds in test_train_opf retightened to ±1% around the new
deterministic point estimates (replacing the wide bounds calibrated
under CUDA non-determinism).

Signed-off-by: Romeo Kienzler <[email protected]>
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Signed-off-by: Romeo Kienzler <[email protected]>
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Signed-off-by: Romeo Kienzler <[email protected]>
@romeokienzler romeokienzler changed the title Bump torch version Bump torch version, make tests deterministic Jul 1, 2026

@albanpuech albanpuech left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is great! I just have a very final request:
Could we launch the regression test with the old pytorch version and see what we get in terms of metrics? That would be to see the impact of bumping the version on the metrics

choices=["true", "warn"],
help=(
"Enable deterministic CUDA/cuDNN algorithms via Lightning Trainer(deterministic=...). "
"Pass --deterministic (alone) for 'warn' mode, or --deterministic true for strict. "

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can you say what warn is in the help please. would be helpful for others

@romeokienzler romeokienzler merged commit 496159c into main Jul 1, 2026
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants