# ref :https://github.com/facebookresearch/xformers/issues/845 import xformers import xformers.ops import torch,os os.environ["MAX_BLOCK_SIZE"] = "655350" os.environ["CUDA_VISIBLE_DEVICES"] = "3" os.environ["CUDA_LAUNCH_BLOCKING"] = "1" q = torch.zeros(([65536, 16, 80])).cuda() k = torch.zeros(([65536, 16, 80])).cuda() v = torch.zeros(([65536, 16, 80])).cuda() out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None) print("Done.")