Post-Training Resource reqiured

[ZZQ0316]

Hello, thank you for sharing this amazing work. What is the minimum computing power requirement for post-training?

[zhengsipeng]

Hello, thank you for sharing this amazing work. What is the minimum computing power requirement for post-training?

Hi, we appreciate your interest in our work.

Our post-training experiments were conducted using a minimum setup of 4 x 80GB GPUs. To maintain training speed, we didn't utilize FSDP sharding or gradient accumulation. However, by enabling these techniques, the model can run on GPUs with 40GB or less VRAM. Could you let us know your GPU resources? We can prioritize corresponding optimizations in future updates. The environment setup and more detailed evaluation code for post-training will be updated soon.

We welcome any further suggestions! Our goal is to contribute to the Embodied AI and VLA community through this open-source project.

Best,
BeingBeyond Team

[ZZQ0316]

Hello, thank you for sharing this amazing work. What is the minimum computing power requirement for post-training?

Hi, we appreciate your interest in our work.

Our post-training experiments were conducted using a minimum setup of 4 x 80GB GPUs. To maintain training speed, we didn't utilize FSDP sharding or gradient accumulation. However, by enabling these techniques, the model can run on GPUs with 40GB or less VRAM. Could you let us know your GPU resources? We can prioritize corresponding optimizations in future updates. The environment setup and more detailed evaluation code for post-training will be updated soon.

We welcome any further suggestions! Our goal is to contribute to the Embodied AI and VLA community through this open-source project.

Best, BeingBeyond Team

Thank you for your reply. We currently only have two A6000 (48GB).