more fixes

changes

Fixes exceeding maximum sequence length when using generate().

Uses native torch decorator for disabling autocast.

Adds disable_autocast support for different device types.

Fixes any potential overflow when calculating attention weights.

Update to new model interface.