Required Training Resources for Reproducing SNav

[2814894078]

Dear authors,

I am working on reproducing the training of SNav (Ours) and replicating the results shown in Table 1 of your paper.

I have a single 48GB GPU. Could you please tell me:

• Is 48GB VRAM enough to fully fine-tune SNav and reproduce the reported results?
• What training precision (BF16/FP16) and batch size did you use?

Thank you very much for your help!

[TidalHarley]

Thank you for your interest in our work.

SNav is a 7B VLN model based on LLaVA-Video-7B-Qwen2. Based on our experience, a single 48GB GPU is likely not sufficient for full fine-tuning to reproduce the reported training results. It is feasible for inference and evaluation, but full fine-tuning usually requires more GPU memory. In our experiment, we train the whole model using 8*A100 80G GPUs.

For training, BF16 is preferred when supported. The effective batch size depends on settings such as sequence length, input resolution, gradient checkpointing, and whether the backbone is fully fine-tuned, so it is difficult to give a universally applicable single-GPU batch size.

That said, our benchmark can be evaluated independently. It is based on HM3D v0.2 [https://github.com/matterport/habitat-matterport-3dresearch] and follows the same data format as VLN-CE [https://github.com/jacobkrantz/VLN-CE], so any model that supports VLN-CE evaluation should be able to evaluate on our benchmark. Usually, you only need to modify the "scene_id" so that it points to the HM3D scenes on your local machine. The whole set of evaluation code has already been released.