debajyotidasgupta/asyncspade-repro-artifacts

Reproduction: AsyncSpade — Efficient Test-Time Scaling with Asynchronous Sparse Decoding

Independent reproduction of ICML 2026 paper #15400 — AsyncSpade: Efficient Test-Time Scaling with Asynchronous Sparse Decoding (Luo, Guan, Li, Wang, Chen; UNC Chapel Hill / JHU / UCLA). arXiv 2510.07486 · OpenReview avVTEzRgba · paper repo UNITES-Lab/AsyncSpade (empty placeholder).

Part of the Hugging Face × AlphaXiv ICML-2026 reproduction challenge.

Why this is an independent reimplementation

The official repo is an empty placeholder (12-byte README, no code) and the method is training-free, so there are no checkpoints to reuse — every result here is reimplemented from the paper's equations (Algorithm 1, Eqs. 1–11) on stock Qwen3 weights.

What reproduces (and what does not)

Claim	Result
C5 temporal locality + linear query prediction (Figs 4-5)	✅ reproduced quantitatively on DeepSeek-R1-0528-Qwen3-8B & Qwen3-32B; recent-16-window overlap 48.9% / 43.9% (paper >40%); ridge one-step-ahead prediction recovers 55–58% of true top-k selection
C2 ≥50% FLOPs reduction vs full attention	✅ exact at FLOPs level: 54.5% (8B, 32k); Fig-9 x-axis reproduced to the digit; accuracy preservation shown at toy scale
C3 cache mgmt fully overlaps inference → optimal TPOT	✅ core ordering reproduced (cache-core < inference 8/8 configs); full optimal-TPOT needs 2-GPU disaggregated serve (out of scope)
C4 cache latency & bandwidth vs inference (Table 2)	✅ pattern reproduced single-GPU; required BW 28–35 GB/s ≪ node 250/350 GB/s; absolute ms differ from 8-GPU node
C1 >20% TPOT vs Quest (Fig 10)	⚠️ mechanism + direction reproduced; shown to be a latency not FLOPs effect; exact magnitude needs the disaggregated serve

Layout

experiments/
  exp_flops.py       exact decode-FLOPs model (Eqs 6-11) — CPU
  exp_mechanism.py   overlap-ratio (Fig4) + ridge query prediction (Fig5) — GPU
  exp_latency.py     inference vs cache-management latency microbench (Table 2) — GPU
  exp_accuracy.py    toy: full vs top-C sparse attention answer agreement — GPU
outputs/
  flops/ mechanism/ latency/ accuracy/   result JSON
  figures/                                Plotly HTML + CSV raw data
make_figures.py      builds the figures from result JSON
build_logbook.py     populates the Trackio logbook claim pages
BRIEFING.md          technical briefing distilled from the paper

Reproduce

pip install torch transformers accelerate datasets plotly huggingface_hub
python experiments/exp_flops.py                                                   # seconds, CPU
python experiments/exp_mechanism.py --model deepseek-ai/DeepSeek-R1-0528-Qwen3-8B --tag qwen3-8b
python experiments/exp_mechanism.py --model Qwen/Qwen3-32B --tag qwen3-32b        # needs 80GB
python experiments/exp_latency.py --tp 8
python experiments/exp_accuracy.py --n-problems 4 --C 1024
python make_figures.py

Compute

HF Jobs was unavailable for this account (402 Payment Required, no credits). All GPU experiments ran on a Vast.ai A100-SXM4-80GB (~$1.04/hr), total ≈ $1.3. The FLOPs model runs on CPU.

Key finding worth flagging for the next agent

The ">20% vs Quest" headline is not a FLOPs claim — at C=2k tokens AsyncSpade and Quest attend to the same tokens and differ by only Quest's tiny 2·q·h·(T/P) page-scan term (FLOPs gap 1–4%). The real saving is wall-clock latency: AsyncSpade removes Quest's sequential-dependent KV selection from the decode critical path by predicting the next query and prefetching asynchronously. So C1 and C3 are fundamentally the same systems result measured two ways, and both require the 2-GPU disaggregated (Inference Rank + Cache Rank) topology to show the full magnitude.