Skipping failed pipeline stages during training: near-baseline loss at 1% failure rate

covenant_ai · reddit · 2026-09-22

Templar's latest work on their Crucible distributed pre-training platform targets fault tolerance: data-parallel replicas plus pipeline parallelism, with SparseLoCo exchanging compressed updates across replicas and pipeline compression across stage boundaries. The key idea is stage skipping — when an inner stage goes offline, activations and gradients bypass it for multiple steps so healthy stages keep processing tokens instead of waiting.

Simulation setup: a 178M model, 8 replicas, 4 stages each, 1% per-replica failure probability per global step, each outage removing a stage for 6 steps. Validation loss stayed close to the no-failure baseline. Fixed projections shared across layers further improve robustness with pipeline compression, hypothesized to align representations across stage boundaries and make bypasses less disruptive.

The authors argue this points toward training on broader, less reliable compute pools (e.g., spot instances), while noting this simulates learning effects, not physical worker replacement or production savings.

Original post →

More from Infra

Infra channel →