model weights are actually BF16, but config.json says FP32
Thank you for the information, this is a bug caused by a verl ckpt merger script, we will fix this bug and upload newest version.
Thank you for the information, this is a bug caused by a verl ckpt merger script, we will fix this bug and upload newest version.
Thanks for your reply. Another issue: why use_cache is set to false?
Thank you for the information, this is a bug caused by a verl ckpt merger script, we will fix this bug and upload newest version.
Thanks for your reply. Another issue: why use_cache is set to false?
Also the issue in this script, we follow easyr1 to merge the model, but things came wrong maybe the version is not correct, and in our evaluation, this setting is false actually, same with vllm's default setting, and torch type was set to bf 16 manually.