4 ms·
Ho thanks for the comment. Verification does not depend on temperature. Expert execution is deterministic (pure matmul). LUMABRI_VERIFY=N re-runs N% of the call
by vforno 2mo ago
Ho thanks for the comment.
Verification does not depend on temperature.
Expert execution is deterministic (pure matmul).
LUMABRI_VERIFY=N re-runs N% of the calls on a second replica and requires byte-identical output.
Temperature (and sampling) happens only on the chatter, after the experts return their activations.
So it can be any value (0, 0.7, 1.2…) without affecting the verification contract.
- stymaar 2mo ago> Expert execution is deterministic (pure matmul). Isn't that only true in theory but wrong in practice due to floating points?
- vforno 2mo ago[dead]
- nikunjbjj 2mo agoThe harder case isn't same-hardware determinism; it's that lumabri wants CPU and GPU peers in the same swarm. Accumulation order differs across hardware, not just across runs.
- snovv_crash 2mo agoThere are fixed-point models too which can be deterministic. But for floating point you are very instruction-set dependent, never mind floating point operations not being commutative.