Reproducible recipe

Precision and numerical gates

Reduce memory or improve throughput without invalidating model math.. Precision choice causes overflow, drift, excess memory, or slow execution.

Status
validated with caveat
Learning path
architecture and kernels
Regression slices
5
Execution
Fresh experiment required

Target

Reduce memory or improve throughput without invalidating model math.

Failure this recipe addresses

Precision choice causes overflow, drift, excess memory, or slow execution.

Data contract

Deterministic numeric fixtures plus one behavioral smoke.

Method or policy

Compare reference and reduced-precision paths with explicit tolerances before measuring speed.

Evaluation contract

Numerical parity and behavioral parity, then memory and throughput.

  • max error
  • loss drift
  • NaN/Inf
  • quality
  • hardware fallback

Budget and stop rule

No long loop until single-step parity passes.

Stop at the first unexplained numerical or behavioral failure.

Decision rule

Adopt the lowest precision that passes correctness and produces a real systems win.

Learning exercise

Measure fp32/fp16 error on a tiny matrix operation and choose a tolerance before looking at the result.

Explain why a faster kernel with wrong numerics is not an optimization.

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.