rl场景使能param_offload,会调用.to执行npu->cpu、cpu->npu的拷贝,发现存在精度问题(训练结果崩溃 梯度爆炸),不使能param_offload精度正常
rl场景使能param_offload精度正常
Atlas A2/CANN 8.5
Thanks for contributing 🎉!
Checklist
🐛 Describe the bug
rl场景使能param_offload,会调用.to执行npu->cpu、cpu->npu的拷贝,发现存在精度问题(训练结果崩溃 梯度爆炸),不使能param_offload精度正常
Expected behavior
rl场景使能param_offload精度正常
Additional context
Environment info
Atlas A2/CANN 8.5
Thanks for contributing 🎉!