Y
Hacker News
new
|
ask
|
show
|
jobs
by
jychang
248 days ago
Tensor parallelism, so you only need to store a fraction of kv cache per gpu.