OmniJev’s Flide serves open decision models through a Jev-style API. It reuses a shared input state when a request asks several typed questions, then runs each question on its own branch.
Share the input, isolate the questions
- A request supplies a state and named questions such as a choice or a yes/no probability.
- For compatible causal models, the engine processes the shared prefix once and forks the question-specific work.
- Its Qwen decoder carries the prefix’s recurrent state into each suffix; attention lets each branch see its own question and the shared prefix.
- Some bidirectional model formats read each question separately, so this sharing does not apply to every supported model.
Check the behavior on your workload
flide verify compares questions asked together with the same questions asked alone. It also checks batches and, when enabled, a session cache. It reports probability differences and bitwise matches; its fork check allows a changed top answer when the reference’s two highest probabilities are within 0.05.
The authors report median latency on one H200 across named workloads. Their own table ranges widely by model and workload; a single maximum is not a general speed guarantee. These measurements were not independently reproduced.
A matching typed answer shows consistency between execution paths, not that the underlying decision is correct or calibrated.
Verification: inspected the API, model-format description, Qwen decoder and verification code at commit 46a9aa32205b393c16c0d13631c98bd20a99f323. The engine and benchmarks were not run; this was not a security audit.