4090d docker install

[lightning1116]

File "/usr/local/lib/python3.12/dist-packages/torch/cuda/init.py", line 478, in _lazy_init
torch._C._cuda_init()
RuntimeError: Unexpected error from cudaGetDeviceCount(). Did you run some cuda functions before calling NumCudaDevices() that might have already set an error? Error 804: forward compatibility was attempted on non supported HW

Driver Version: 570.211.01 CUDA Version: 12.8 is not enough?

[foreverlms]

Thanks for the report, we haven't tested on 4090 yet, so could you please describe this more clearly: Is this the error of docker install that you titled or when running the inference?

I think it is better to paste the reproduce steps.

[junichi-2099]

关于 DROID 数据集的那个 Issue，我也挺好奇 Cosmos 处理大规模机器人数据的策略。World Model 做预训练虽然能学到物理规律，但在真实机器人上落地时，动作空间的映射往往比视觉生成更棘手。特别是那个 Cosmos3-Nano-Policy，在处理长序列操作时，怎么平衡推理速度和物理一致性？如果单纯靠视频生成做开环预测，反馈延迟这一块在真实场景会不会是个大坑？

[foreverlms]

关于 DROID 数据集的那个 Issue，我也挺好奇 Cosmos 处理大规模机器人数据的策略。World Model 做预训练虽然能学到物理规律，但在真实机器人上落地时，动作空间的映射往往比视觉生成更棘手。特别是那个 Cosmos3-Nano-Policy，在处理长序列操作时，怎么平衡推理速度和物理一致性？如果单纯靠视频生成做开环预测，反馈延迟这一块在真实场景会不会是个大坑？

Thanks for this question. Did you mean #207 ? If so, please paste your question in English under the related issues, I think this is not related with this issue. Though I could understand what you mean, but it's better in English for colleagues who are looped to. I focus mainly on training infra side.