Reproducible recipe

StreamingLLM and KIVI cache compression

Bound or compress KV memory for long local contexts.. KV cache growth, not model weights, is the measured memory blocker.

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

Target

Bound or compress KV memory for long local contexts.

Failure this recipe addresses

KV cache growth, not model weights, is the measured memory blocker.

Data contract

Long-context replay prompts with cached-forward quality checks.

Method or policy

Retain attention sinks/window tokens or quantize KV groups, then replay identical continuations.

Evaluation contract

Cache bytes and trajectory/perplexity parity versus full KV.

  • long-context recall
  • greedy prefix match
  • throughput
  • storage packing
  • composition

Budget and stop rule

Existing smokes are closed; new work requires a long-context product blocker.

Stop if memory savings lack a packed representation or quality gate.

Decision rule

Keep as an opt-in long-context lever until quality and composition gates pass.

Learning exercise

Calculate full, windowed, int8, and int4 KV bytes for one model/context.

Choose between eviction and quantization from the actual context failure mode.

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.