Comment by dist-epoch
4 days ago
After you load the weights into the GPU and keep the KV cache there too, you don't need any other significant traffic.
4 days ago
After you load the weights into the GPU and keep the KV cache there too, you don't need any other significant traffic.
Even in tensor parallel modes? I thought it could only work if you're fine stalling all but n GPU for n users at any given moments.