Preprint / scholarly article
Certified Bottleneck Floors for Transformer Training
- Published
- DOI
- 10.5281/zenodo.18042480
Abstract
The preprint defines telemetry contracts that certify conservative lower bounds on required data movement, checkpointing, and collective communication for transformer training, yielding fail-closed bottleneck floors and actionable slack intervals without vendor-specific instrumentation.
Keywords
- transformer training
- bottleneck floors
- data movement
- checkpointing
- collective communication
- all-reduce
- telemetry contract
- auditability
- throughput ceiling
- I/O complexity
- fail-closed certificates