* Fix IPEX attn_drop * IPEX SDPA 4GB fix Alchemist GPUs lack the 64 bit emulation needed for >4GB in this case. Attention slicing workaround copied from https://github.com/vladmandic/automatic/blob/master/modules/intel/ipex/attention.py * Do 4GB fix only when necessary ~20% performance boost to not use the 4GB fix when not necessary * Move attention.py inside utils * Better 4GB estimation * Remove unnecessary warning