Skip to content

[Fix] Fix TP sampler inconsistency bug#85

Open
DarkSharpness wants to merge 1 commit intosgl-project:mainfrom
DarkSharpness:fix_sampler
Open

[Fix] Fix TP sampler inconsistency bug#85
DarkSharpness wants to merge 1 commit intosgl-project:mainfrom
DarkSharpness:fix_sampler

Conversation

@DarkSharpness
Copy link
Collaborator

@DarkSharpness DarkSharpness commented Feb 26, 2026

For the bitwise equal input, the flashinfer sampler gives inconsistent output token ids for different ranks. To guarantee consistency, we force to broadcast the result from rank 0. fix #84

@DarkSharpness
Copy link
Collaborator Author

DarkSharpness commented Feb 26, 2026

This should fix #84 . The bug is from flashinfer side. I will open a issue later there with a minimal reproduction example. The root cause: I forgot to manual set the random seed for our engine.

@jiahe7ay
Copy link
Contributor

@DarkSharpness I've tested the same inputs multiple times and also tried different ones. The issue is gone—it seems to be fixed

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.

[BUG] GPU Memory leak during multi-GPU inference.

2 participants