x-small / benchmarks /tput_release.csv
kenpeter123's picture
Add benchmark graphs: perplexity trend and 12-layer throughput
3d101dd
Raw History Blame Contribute Delete
242 Bytes
batch,seq,tps,vram_gb,mode
1,2048,14892,2.04,"eager fwd+bwd, grad-ckpt on, bf16"
2,2048,21625,3.63,"eager fwd+bwd, grad-ckpt on, bf16"
4,2048,22720,6.75,"eager fwd+bwd, grad-ckpt on, bf16"
8,2048,OOM,,"eager fwd+bwd, grad-ckpt on, bf16"