Floating point exception when running the official code

[TsingZ0]

When running the official code from HuggingFace, an error occurs: Floating point exception (core dumped).

GPU Machine: H20
Torch: 2.2.0
transformers: 4.51.3

[TsingZ0]

The official code uses bf16, but it failed.

After changing to fp32, it succeeds.

[TsingZ0]

Updating nvidia-cublas-cu12==12.1.3.1 to nvidia-cublas-cu12 12.4.5.8 solves this issue.