DAILY BRIEF / KERNELS
FlashInfer 0.7.4 adds batch-invariant kernels.
What changed
Selected GEMMs and quantizers can produce the same output-row bits across batch sizes on B200, B300, and Rubin GPUs.
Why it matters
This is useful to investigate when numerical agreement changes with batching. Validate the complete model path separately.
Scope and limitations
The flag requires backend="auto". Unsupported inputs or devices raise an error; this is not a blanket determinism guarantee for every operator.
Go to the source
A sourced technical brief published 16 Nov 2026. The implication is keplerfold’s interpretation. We have not independently benchmarked this change.