3 ms·
For inference the bandwidth is generally not parallelized because the weights need to go through the model layer by layer. The most common model splitting metho
by choilive 2y ago
For inference the bandwidth is generally not parallelized because the weights need to go through the model layer by layer. The most common model splitting method is done by assigning each GPU a subset of the LLM layers and it doesn't take much bandwidth to send model weights via PCIE to the next GPU.
- manmal 2y agoMy understanding is that the GPU must still load its assigned layer from VRAM into registers and L2 cache for every token, because those aren’t large enough to hold a significant portion. So naively, for a 24GB layer, you‘d need to move up to 24GB for every token.