Whitepapers

A living archive of Stillpoint experiments. Each paper owns one technical topic and accumulates its upstream context, reproducible lab setup, methods, measured results, failures, limitations, artifacts and subsequent runs.

Measured results only. Every number must identify its hardware, runtime, workload and repetition count.

PIN-0001
GREEN · 12/12

Persistent standing-place restore for hybrid GDN models

Exact LMCache restore across a full vLLM restart, including the LMCache 0.5.3 root causes, patch, verification suite and measured restore speedups.

  • 97.0–98.8% external cache hit
  • Up to 7.6× lower end-to-end wall time
  • Exact buried-key recovery after restart
Read summary →PDF ↓
PIN-0005
VALIDATED PROFILE · 40/40

P-PAS+ adaptive scheduling with persistent Stillpoint restore

Porting prefill-pressure adaptive scheduling to the complete Qwen3.8 production profile without losing LMCache, MTP-3, long context or serving capabilities.

  • All candidate equivalence and restore gates passed
  • 40/40 repeated candidate workloads completed
  • +4.90% paired decode throughput in the target pressure case
Read summary →PDF ↓

Standard paper structure

  1. Abstract and current status
  2. Problem and motivation
  3. Prior work and author credit
  4. System architecture and constraints
  5. Lab hardware and software setup
  6. Implementation or adaptation
  7. Methodology and acceptance gates
  8. Measured results
  9. Failures and negative results
  10. Limitations and non-claims
  11. Artifacts and reproducibility
  12. Run history and next experiments