WiP: Efficient Speculative Decoding for AI PCs via Hierarchical N-Gram Retrieval
Bibliographic record
Abstract
AI PCs are emerging as a promising platform for on-device LLM inference, but their limited memory and compute resources pose significant challenges to Large Language Model (LLM) inference, which underutilizes available computing resources in decoding. Speculative decoding addresses this by generating multiple draft tokens and verifying them in parallel using the target model, improving efficiency over purely autoregressive methods. However, existing speculative decoding methods struggle on AI PCs: model-based approaches are bottlenecked by non-negligible draft model inference time, while retrieval-based methods face a trade-off between insufficient retrieval quality and excessive memory usage from external datasets. To address this fundamental trade-off, we propose NG+, a work-in-progress retrieval-based LLM inference system optimized for AI PCs. NG+ is designed to deliver retrieval quality of large-scale datasets within constrained memory budgets by employing a hierarchical n-gram caching architecture---storing frequent n-grams in memory and offloading less frequent ones to SSD. In addition, NG+ overlaps CPU-based n-gram retrieval for the next decoding step with iGPU-based verification of the current draft, effectively prefetching from SSD to hide I/O latency. Preliminary evaluations on an AMD Ryzen™ AI MAX+ 395 platform demonstrate NG+'s efficacy, achieving a 1.62x speedup over auto-regressive decoding, outperforming existing baselines.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.006 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.004 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.006 | 0.006 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".