Intel GPUs & 16-bit training support

[ThanatosShinji]

Intel Arc 770 ubuntu 24.04， pytorch==2.8.0+xpu
Train 8-block RRDBNet, gt_size: 512, dtype=bfloat16

cmd:

export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
export OverrideDefaultFP64Settings=1
export IGC_EnableDPEmulation=1
export ExperimentalCopyThroughLock=0
export NEOReadDebugKeys=1
python -m torch.distributed.launch --nproc_per_node=2 --master_port=4321 basicsr/train.py -opt options/train/RealESRGAN/train_realesrnet_x4plus.yml --launcher pytorch

2 GPUs

2025-10-30 12:20:37,309 INFO: [train..][epoch:  0, iter:   1,400, lr:(2.000e-04,)] [eta: 13 days, 20:41:31, time (data): 1.184 (0.004)] l_pix: 7.8613e-02 
2025-10-30 12:22:35,218 INFO: [train..][epoch:  0, iter:   1,500, lr:(2.000e-04,)] [eta: 13 days, 20:17:01, time (data): 1.175 (0.004)] l_pix: 7.6172e-02 
2025-10-30 12:24:33,601 INFO: [train..][epoch:  0, iter:   1,600, lr:(2.000e-04,)] [eta: 13 days, 20:00:17, time (data): 1.180 (0.004)] l_pix: 7.9102e-02 
2025-10-30 12:26:33,508 INFO: [train..][epoch:  0, iter:   1,700, lr:(2.000e-04,)] [eta: 13 days, 20:00:10, time (data): 1.184 (0.004)] l_pix: 6.4453e-02 

Single GPU

2025-10-29 13:06:40,548 INFO: [train..][epoch:  0, iter:   1,800, lr:(2.000e-04,)] [eta: 11 days, 14:28:19, time (data): 1.011 (0.004)] l_pix: 8.5938e-02 
2025-10-29 13:08:25,837 INFO: [train..][epoch:  0, iter:   1,900, lr:(2.000e-04,)] [eta: 11 days, 15:09:09, time (data): 1.038 (0.004)] l_pix: 6.1035e-02 
2025-10-29 13:10:05,298 INFO: [train..][epoch:  0, iter:   2,000, lr:(2.000e-04,)] [eta: 11 days, 14:57:18, time (data): 1.015 (0.004)] l_pix: 8.1055e-02 

3090
single

2025-10-30 22:42:11,037 INFO: [train..][epoch:  5, iter:     700, lr:(2.000e-04,)] [eta: 5 days, 6:34:58, time (data): 0.460 (0.012)] l_pix: 6.2842e-02

dual

2025-10-30 22:47:44,661 INFO: [train..][epoch:  4, iter:     300, lr:(2.000e-04,)] [eta: 5 days, 22:15:47, time (data): 0.504 (0.051)] l_pix: 7.6922e-02