1. Attach the inf-check hooks if the grad scale is getting too small. 2. Add try-catch to avoid OOM in the inf-check hooks. 3. Set warmup_start=0.1 to reduce chances of divergence
1. Attach the inf-check hooks if the grad scale is getting too small. 2. Add try-catch to avoid OOM in the inf-check hooks. 3. Set warmup_start=0.1 to reduce chances of divergence