A high-performance inference engine for AI models
Benchmark instructions
Code Review Rules
Benchmark model parity
These files define the same benchmark request and response models. Paths are relative to this directory:
common-cpp/src/bench.hppcommon-py/src/bench.pyengine-mlxserve/src/bench.zigengine-uzu/src/bench.rs
When changing or reviewing any of these files, read all four, including files
absent from the diff. Keep ChatMessage, BenchSampling, BenchRequest,
BenchResponse, and any new shared models consistent across the implementations.
- Flag additions, removals, or renames of shared models or fields that are not reflected in the other implementations.
- Compare equivalent field types, nested models, collection element types, optionality, defaults, and serialized field names. Account for each language's native types rather than requiring identical syntax.
- Check changes to numeric ranges, enum constraints, and omission versus JSON
nullfor incompatible behavior across implementations. - Helper methods, properties such as Python's
BenchRequest.prompt, derives, and language-specific implementation details do not need counterparts unless they change the serialized model or accepted inputs. - Report mismatches introduced or worsened by the reviewed change, identifying the model, field, and counterpart files that need updating. Do not report an existing difference as a new regression.