Skip to content

extending.md

Latest commit

 

History

History
27 lines (16 loc) · 3.06 KB

extending.md

File metadata and controls

27 lines (16 loc) · 3.06 KB
Raw

Worked contributor examples

Add a configuration option

Suppose a future change adds a limit to a selection stage. First find the owning policy (RuntimeCandidateSelectionConfig and its service), and decide whether the option actually affects behavior. Do not add another dormant knob.

  1. Add a clearly named typed property with its default and a comment describing the decision it controls.
  2. Add the underscored YAML key to config.default.yaml; decide explicitly whether the distributed tuning profile uses the same default.
  3. If a CLI override is useful, add it to MagicQuantYamlLoader.ApplyCliOverrides and the value/flag contract in CliOptionValidator. Validate numeric input with invariant culture. Add preflight validation for constraints that should fail before work begins.
  4. Add tests for omitted/default values, CLI precedence, and the observable stage behavior. Keep any changed Config.Current/Cache state scoped and restored.
  5. Update command help and configuration documentation. Run the strict distributed-config tests and the complete suite.

YamlConfigurationDiagnostics derives known keys from the typed schema, so it does not need a duplicate property-name list. Free-form readme.frontmatter and dictionary keys remain user-defined.

Change a path or process invocation

For a new benchmark option, update BenchmarkCommands, then assert the literal argv sequence in BenchmarkContractTests. Use NativeCommand.CreateStartInfo, not interpolated shell strings. Include a path with spaces and metacharacters in the test. Keep retries in the calling service; ProcessRunner returns a nonzero exit code and only throws for launch/IO/cancellation failures.

For conversion behavior, NativeModelConversionService accepts IProcessRunner. NativeConversionTests injects a small fake that writes a partial output and returns failure or cancellation. That verifies incomplete artifacts never acquire a reusable success marker without invoking Python or a model. Keep production code using the real runner by default.

For output destinations, update OutputPathService and PathSafety, preserving existing command semantics or documenting a deliberate migration. Test ordinary paths, parent/child collisions, similarly prefixed sibling directories, and linked directories. Run the optional smoke workflow if native argument or artifact lifecycle behavior changed.

Work on numerical policy

Read the research wiki and the owning service before editing. Add a regression around the actual measured/predicted tradeoff and its context identity. Do not replace exact custom tensor assignments with a built-in family surrogate merely to make a test pass. Model hash, architecture/profile identity, and imatrix scope are part of the input.

Global configuration and registries still exist; this cleanup does not support multiple concurrent campaigns in one process. New helpers should accept explicit inputs, return results, and be testable without mutating those registries. Extract a responsibility with behavior tests rather than mechanically splitting a large class into partial files.