Stream the checkpoints

Fabric 09
Data centerHostHostHostHostTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPTCPGPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…GPU (H1…NVLink…NVLink…NVLink…NVLink…InfiniB…200 nsInfiniB…200 nsInfiniB…200 nsInfiniB…200 nsInfiniB…200 nsInfiniB…200 nsInfiniB…200 nsInfiniB…200 ns

Fix the board. Turn every goal green.

Goals

Not tested yet

The share of requests that get a correct answer.How much recent data a failure can lose, in time.What the servers and services cost each month.

32 H100 train a 13B model. Every 20 steps all ranks save 182 GB of weights and optimizer state, and the checkpoint store takes 10 GB/s. A GPU may fail at any time. Keep the GPUs busy 95 percent of the run and lose at most 300 s of work in a failure. One change.

11:40 Goodput 87 percent. Graph: GPU use drops to 0 for 18 s every 140 s, store write flat at 10 GB/s.

A disk that fillsArrow keys change the write rate. Tap to ship the logs.

5 MB/s · logs kept on the host · disk full in 33 minutes