Skip to content

replication.md

Latest commit

 

History

History
30 lines (23 loc) · 1.21 KB

replication.md

File metadata and controls

30 lines (23 loc) · 1.21 KB
Raw

Replicate the paper

The paper branch preserves the benchmark implementation, including research-only selectors, KV transforms, capture-bank replay and their original tests; these are outside the public API on main. The paper-v1 tag fixes the first replication snapshot.

In a separate checkout, select the snapshot:

git switch --detach paper-v1

Follow that snapshot's README for installation and the exact benchmark commands. It includes:

  • Twelve configs for FanOutQA and LongBench v2, including the request ablation.
  • Qwen 3, Gemma 4, Ministral 3, and Nemotron serving runtimes.
  • A requirements file for each of the two measured vLLM stacks.
  • The natural FanOutQA and selector-development archives under data/.
  • CPU tests, result scoring, and handoff-recall analysis.

The bundled archives unpack offline with digest verification. Model weights and the pinned LongBench dataset download separately. Gemma and Ministral require accepted model access terms. The registered benchmark runs use one node with eight H100 80 GB GPUs; the API's small-model example does not reproduce their accuracy or latency.

The manuscript is Receiver-Conditioned Latent Communication gives 94% CacheBack. Paper link: TBA.