Skip to content

commands.md

Latest commit

 

History

History
74 lines (50 loc) · 5.14 KB

commands.md

File metadata and controls

74 lines (50 loc) · 5.14 KB
Raw

Commands and workflows

Run these examples with the NuGet-installed magicquant CLI. Replace paths and identities with your own. For source builds, substitute dotnet run --project src/MagicQuant -c Release --no-build -- for magicquant. mq in older help is shorthand, not an installed command.

Config creation and version

magicquant init-config --output config.yaml copies the bundled tuning profile without runtime setup and refuses to overwrite a file. magicquant --version prints the application version and available source revision.

Full discovery pipeline

magicquant pipeline \
  --config config.local.yaml \
  --model-dir /data/models/my-model \
  --architecture-family my-model-family \
  --output-dir /data/exports/my-model \
  --output-name-prefix MyModel

The pipeline converts/loads the native source, reviews tensor groups, resolves identity and baselines, measures isolation samples, predicts useful hybrids, validates them, and exports survivors. The exact runtime choices depend on benchmark truth and YAML policy. evolution invokes the same implementation. build-hybrids also enters the full pipeline; it is not limited to exporting previously selected files.

--use-imatrix enables the configured acquisition/build flow. Provide an appropriate imatrix source in YAML. --recheck-hardware-probe refreshes execution-plan probing after hardware changes. --skip-tensor-group-confirm suppresses the tensor-group prompt for an already reviewed unattended campaign; it does not bypass all other confirmations.

Clone known tensor configurations

For the complete publisher-to-user walkthrough, including manifest filenames and pointing at another Hugging Face release, see manifests and cloning.

magicquant clone-repository-quants \
  --config config.local.yaml \
  --model-dir /data/models/compatible-model \
  --architecture-family my-model-family \
  --source-json /data/releases/source/magicquant-manifest/magicquant.clone-configs.json \
  --output-dir /data/exports/cloned-model

Use --source-repo owner/repo instead to read a Hugging Face repository. --source-json also accepts an HTTP(S) URL. Clone mode rebuilds the tensor configurations and benchmarks them locally; it does not establish that the new model passed full discovery.

Clone mode rebuilds all entries present in the manifest, including configurations originally learned from external providers. It does not use output.export_external_learned_baselines, which controls the pipeline's choice between local exports and upstream links. This is useful when cloning to a fine-tuned or uncensored variant that the original provider does not host. Review the upstream-link versus variant-export guidance and limits of tensor-configuration copying.

By default the manifest must match the target tensor inventory. --allow-missing-manifest-tensors explicitly allows a strict subset; unmatched target tensors use base quantization. --missing-manifest-base-quant Q8_0 additionally selects that base quant. Use these only when that compatibility tradeoff is intended.

Validate predictions against existing measurements

magicquant validate-predictions \
  --config config.local.yaml \
  --model-dir /data/models/my-model \
  --architecture-family my-model-family \
  --output-dir /data/reports/my-model

This writes prediction_validation_general.csv and prediction_validation_general.md. It uses existing general-category SQLite benchmark truth rather than launching a fresh full discovery campaign. Startup still runs the common dependency validation and stale-artifact cleanup.

For imatrix measurements supply --imatrix-path /data/imatrix.dat or --imatrix-identity-hash <sha256> to select the exact bucket. Merely enabling flags.use_imatrix does not select a validation bucket. Without either option, validation uses no-imatrix truth.

Rerun and reuse

magicquant pipeline \
  --config config.local.yaml --reuse-existing-final-artifacts

The normal pipeline can reuse scoped measurements, but final export normally cleans/rebuilds output. Artifact reuse is an additional opt-in: pipeline exports require exact filename and benchmark byte-size matches; clone mode also validates its benchmark JSON rows. A file merely existing is not sufficient.

Help and exit status

No arguments, help, --help, or -h display top-level help. <command> --help and <command> -h display command help without config loading, cleanup, database access, or dependency installation.

The host returns 0 on normal completion/help, 2 for an unknown command, 130 for cooperative cancellation, and 1 for an exception caught at the command boundary. Services may handle individual candidate failures and continue a campaign, so also inspect the reported sample failures and final artifacts.

Use --check-config on a normal command to validate local inputs without running it. Use --strict-config to reject unknown/inactive YAML settings rather than warning. See testing for the opt-in real-model workflow.