Zhang JiaXin

icon indicating an email [email protected]

icon location XiaMen

Results 4 comments of Zhang JiaXin

[BUG] RuntimeError: inflight params error when using DeepSpeed for Reinforcement Learning

I have the same problem @HeyangQin

[BUG] RuntimeError: still have inflight params [<bound method Init._convert_to_deepspeed_param.<locals>.ds_summary of Parameter containing:

@HeyangQin I also met this problem when running deepspeed chat the deepspeed version is 0.9.5

[BUG] RuntimeError: still have inflight params [<bound method Init._convert_to_deepspeed_param.<locals>.ds_summary of Parameter containing:

> Hello @ZJXNEFU. Could you provide a reproduce script for us to investigate this issue? Thank you ``` # deepspeed-chat step3 single_node run_rl.sh #!/bin/bash ACTOR_MODEL_PATH=/model/path # a model fine-tuned on...

Much more memory used in step 3 when using multi gpus compared to using single gpu

Same hardware environment，Same problem. I just select a model that has 15B parameters as Actor, but how does the 30B-opt model work properly?