Hi,
Thank you for the paper and the code.
I have a question about the block-size scaling experiment in TraceRL. In the paper, you mention that after 60 steps, rollout was also switched to block size 8.
When using block size 8, did you also change denoising_steps_per_block to 8, or did you keep the original denoising step setting (4) while only changing the block size?
I am also confused about the static evaluation setting, since in my reproduction it becomes extremely slow with block size 32 for SDAR.
Thanks in advance.
Hi,
Thank you for the paper and the code.
I have a question about the block-size scaling experiment in TraceRL. In the paper, you mention that after 60 steps, rollout was also switched to block size 8.
When using block size 8, did you also change
denoising_steps_per_blockto 8, or did you keep the original denoising step setting (4) while only changing the block size?I am also confused about the static evaluation setting, since in my reproduction it becomes extremely slow with block size 32 for SDAR.
Thanks in advance.