Hi authors,Thank you for your impressive work on Flash-VStream!
I am very interested in your two-process framework and the Flash Memory design.I am planning to replicate the training results or fine-tune the model on my own hardware. In your paper, you mentioned that the training was conducted using 8 x NVIDIA A100 (80GB) GPUs with a total batch size of 64.
To better evaluate the hardware requirements for my setup, could you please provide some details regarding the peak GPU memory usage (VRAM) per card during the training/instruction tuning stage?
Specifically, I would like to know:
Under your default configuration (LoRA fine-tuning, BFloat16, DeepSpeed Stage 2, and NVtokens = 11520), approximately how many GBs of VRAM were utilized on each A100?
Do you have any suggestions or minimum VRAM requirements for training on GPUs with smaller memory capacity (e.g., 48GB)?
Looking forward to your reply. Best regards!
Hi authors,Thank you for your impressive work on Flash-VStream!
I am very interested in your two-process framework and the Flash Memory design.I am planning to replicate the training results or fine-tune the model on my own hardware. In your paper, you mentioned that the training was conducted using 8 x NVIDIA A100 (80GB) GPUs with a total batch size of 64.
To better evaluate the hardware requirements for my setup, could you please provide some details regarding the peak GPU memory usage (VRAM) per card during the training/instruction tuning stage?
Specifically, I would like to know:
Under your default configuration (LoRA fine-tuning, BFloat16, DeepSpeed Stage 2, and NVtokens = 11520), approximately how many GBs of VRAM were utilized on each A100?
Do you have any suggestions or minimum VRAM requirements for training on GPUs with smaller memory capacity (e.g., 48GB)?
Looking forward to your reply. Best regards!