Skip to content

Use the Python Inductor wrapper for Cosmos3 Nano 80 GB image presets - #3413

Merged
bghira merged 1 commit into
mainfrom
fix/issue-3352-dynamo-wrapper
Nov 21, 2026
Merged

bghira merged 1 commit into
mainfrom
fix/issue-3352-dynamo-wrapper

Conversation

@bghira

@bghira bghira commented Nov 21, 2026

Copy link
Copy Markdown
Owner

Closes #3352.

Use Inductor's Python wrapper for the standard and legacy Cosmos3 Nano 80 GB image recipes. On H100, the C++ wrapper reproducibly produced a non-finite Q projection in generator block 1 with either FlashAttention 3 or Diffusers attention. The Python wrapper preserves regional Inductor compilation and the original attention choice.

Validation

  • Both 80 GB recipes completed 100 training steps with crop disabled on a single H100 80 GB with this change.
  • The 48 GB preset completed 100 steps on both current main and the original sweep revision.
  • Passed 22 Dynamo and cache tests.

@bghira
bghira merged commit 2e17790 into main Nov 21, 2026
2 checks passed
@bghira
bghira deleted the fix/issue-3352-dynamo-wrapper branch November 21, 2026 20:39
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.

Cosmos3 Nano 48 GB and 80 GB image presets produce a non-finite loss on the first batch

1 participant