Reproducible recipe

Speculative heads and decoding

Increase local decode throughput after model quality is accepted.. Autoregressive decode latency is the measured user-visible bottleneck.

Status
validated with caveat
Learning path
runtime and agents
Regression slices
5
Execution
Fresh experiment required

Target

Increase local decode throughput after model quality is accepted.

Failure this recipe addresses

Autoregressive decode latency is the measured user-visible bottleneck.

Data contract

Target-domain sequences for head training plus a frozen acceptance and quality suite.

Method or policy

Train draft heads or a draft model, verify proposed tokens with the target model, and measure accepted tokens per unit work.

Evaluation contract

End-to-end tok/s at output parity.

  • acceptance rate
  • quality parity
  • memory
  • training cost
  • short prompts

Budget and stop rule

Medusa/EAGLE smoke retained; a production recipe is a new training experiment.

Stop if acceptance cannot repay verification and memory overhead.

Decision rule

Adopt only on an end-to-end decode benchmark, not head acceptance alone.

Learning exercise

Compute expected speedup from draft length, acceptance rate, and verification cost.

Identify when speculative decoding becomes slower than ordinary decoding.

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.