Skip to content

fix(gms): rebind non-parameter tensors after publish - #11279

Merged
dagil-nvidia merged 2 commits into
release/1.3.0from
schwinns/cherrypick-gms-publisher-rebind-1.3.0
Jul 7, 2026
Merged

fix(gms): rebind non-parameter tensors after publish#11279
dagil-nvidia merged 2 commits into
release/1.3.0from
schwinns/cherrypick-gms-publisher-rebind-1.3.0

Conversation

@galletas1712

@galletas1712 galletas1712 commented Jul 6, 2026

Copy link
Copy Markdown
Contributor

Summary

Cherry-pick #11201 to release/1.3.0 to prevent fatal GPU write faults when a GMS publisher writes mutable non-parameter tensors after the weights mapping becomes read-only.

  • Rebinds GMS-resident buffers and tensor attributes to private CUDA allocations after publish, while parameters remain on the shared read-only mapping.
  • Preserves importer materialization and GMS metadata behavior.
  • Includes the end-to-end regression test from the merged main PR.
  • Clean cherry-pick with no content changes or merge conflicts.

Original PR: #11201
Source merge commit: 56a74283f705166a1f8bf48f1306f42d6846287c
Release commit: f11595767389d52ec6c4fad977674569bde1479f

Dependencies

None. The source PR is merged to main, and this change has no stated PR dependency.

Validation

  • Fresh release-branch CI at e0cefb302da6db1e4724871f2f8f0efbcb1bdd19: 58 passed, 36 expected skips, 0 failed or pending.

  • Exact GPU job dynamo-runtime / test / gpu cuda13.0, amd64 ran test_finalize_gms_write_rebinds_nonparameter_tensors: passed; suite summary 28 passed, 3 skipped, 0 failed.

  • Added focused list/tuple tensor coverage; pre-commit and DCO checks passed.

  • Stable patch ID matches fix(gms): rebind non-parameter tensors to private memory after publish #11201: 6a2263e8bf746cda479e6ba5f2de8e7d10a0293b.

  • pre-commit run --files lib/gpu_memory_service/client/torch/module.py lib/gpu_memory_service/integrations/common/utils.py lib/gpu_memory_service/tests/test_torch_integration.py — passed.

  • git diff --check origin/release/1.3.0...HEAD — passed.

  • DCO signoff is present on the release commit.

  • The original PR passed its live-GMS GPU regression suite (5 passed), full pre-commit suite, and stack-level snapshot restore E2E. The GPU/live-GMS test could not be rerun locally; release-branch CI executed it successfully at the tested head.

#11201)

Signed-off-by: Schwinn Saereesitthipitak <schwinns@nvidia.com>
@galletas1712
galletas1712 requested review from a team as code owners July 6, 2026 21:23
@github-actions github-actions Bot added the fix label Jul 6, 2026
@galletas1712 galletas1712 changed the title fix(gms): [cherry-pick 1.3.0] rebind non-parameter tensors after publish fix(gms): rebind non-parameter tensors after publish Jul 6, 2026
@datadog-official

This comment has been minimized.

Signed-off-by: Dan Gil <dagil@nvidia.com>

@dagil-nvidia dagil-nvidia 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.

Release backport quality gate: exact source patch plus focused list/tuple rebinding coverage; the named live-GMS regression passed on GPU and the full fresh rollup completed 58 passed / 36 expected skips / 0 failures at head e0cefb3.

@dagil-nvidia
dagil-nvidia merged commit 5295421 into release/1.3.0 Jul 7, 2026
95 checks passed
@dagil-nvidia
dagil-nvidia deleted the schwinns/cherrypick-gms-publisher-rebind-1.3.0 branch July 7, 2026 02:08
rst0git pushed a commit to doublewordai/dynamo that referenced this pull request Jul 14, 2026
Signed-off-by: Schwinn Saereesitthipitak <schwinns@nvidia.com>
Signed-off-by: Dan Gil <dagil@nvidia.com>
Co-authored-by: Dan Gil <dagil@nvidia.com>
(cherry picked from commit 5295421)
Signed-off-by: Radostin Stoyanov <radostin@doubleword.ai>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants