torch.AcceleratorError: CUDA error: device-side assert triggered
#14
by
TurboPascal - opened
在transformers更新5.0之后训练报错 目前定位在这里
rope_cos = rope_cos[position_ids_idx].unsqueeze(2) # [bs, seq_len, 1, dim]
rope_sin = rope_sin[position_ids_idx].unsqueeze(2) # [bs, seq_len, 1, dim]