Reproducible recipe

Quantization and storage formats

Reduce model download, RAM, or execution cost on Mac and browser runtimes.. Full precision exceeds the target's distribution or memory budget.

Status
validated with caveat
Learning path
quantization and packaging
Regression slices
5
Execution
Fresh experiment required

Target

Reduce model download, RAM, or execution cost on Mac and browser runtimes.

Failure this recipe addresses

Full precision exceeds the target's distribution or memory budget.

Data contract

Calibration samples plus frozen quality, load, and runtime gates.

Method or policy

Apply the selected format, verify roundtrip/load parity, then measure actual kernel/runtime behavior.

Evaluation contract

Task quality and perplexity at measured file size, RAM, and tok/s.

  • rare tokens
  • long context
  • load parity
  • cold start
  • kernel support

Budget and stop rule

Storage smokes are retained; large-model conversion or calibration is a fresh bounded experiment.

Stop if a smaller file does not load, preserve quality, or reach a matching runtime kernel.

Decision rule

Name storage-only quantization honestly; claim runtime gains only with measured kernels.

Learning exercise

Trace one fp16-to-int4 artifact from calibration through load and decode measurement.

Explain why four-bit storage can still execute with wider arithmetic.

How to read this dossier

This canonical page separates the retained repository decision from the material that informed it. A source may describe an external claim, historical measurement, local observation, or planned procedure; those evidence classes are not interchangeable. Follow the provenance links before reusing a number or method.

Status and disposition describe what PostTrainLLM retained when this record was indexed. They are not a live product promise, a newly run benchmark, or permission to restart historical work. Unknown or unmeasured fields remain unknown rather than being treated as zero.

To reuse the record, first name the exact claim you need and trace it to the linked source. Then check whether the original environment, model revision, data split, evaluator, hardware, and budget match the proposed use. If they do not, treat the record as a hypothesis or design reference and run the smallest fresh comparison that can falsify it. Preserve negative outcomes and regressions beside any improvement; a local win on one slice does not silently become a general capability claim.

Source provenance

The normalized record comes from docs/recipes/registry.json. The links below are the tracked evidence and explanatory sources preserved with the record.