timesteps bug

[byliutao]

in flow_grpo/diffusers_patch/pipeline_with_logprob.py
Here, to ensure precision and align with the SD3 or Flux pipeline, should this be changed to torch.float

[zhengkw18]

Actually, I don't think this is a bug. The timesteps of the models are distributed in [0,1000]. Using INT representation will result in very small quantization error. I think this is a standard practice in sampling pipelines of these models.