Exact and approximate modes
Distribution and approximation are separate choices in Mycellios.
Exact mode
The reference-equivalent path changes where the model is computed without silently changing the graph it must execute. It is the default contract for features such as:
- contiguous stage partitioning;
- local KV cache per stage and request;
- identity-preserving transport;
- exact speculative verification with rollback;
- recovery to a compatible stage or route.
Bit identity can still depend on the backend, dtype, kernel, and collective implementation. A deployment must run its own parity canary rather than infer equivalence from configuration alone.
Approximate mode
Compression, reduced-precision activation codecs, omitted work, or a changed model graph may improve latency or memory use, but they can also change the generated tokens.
An approximate optimization must be:
- explicitly enabled;
- labeled in the route or request policy;
- measured against an agreed reference;
- removable without changing the default contract.
Current analytical or experimental codecs are not a public performance promise. Use only modes validated for the deployed model, route, and acceptance threshold.
Practical rule
If an optimization changes the answer the model may produce, it is not merely a transport detail. Treat it as a model-quality decision and expose it to the operator.