thc1006/qwen3.8-speculative-decoding-rtx3090 is a pre-registered experiment repository comparing speculative decoding for Qwen3.8-27B on a single RTX 3090. The repository covers llama.cpp-based experiments and records that, with the built-in draft method draft-mtp set to --spec-draft-n-max 2, the server-reported decoding throughput was 59.8% higher across 25 inputs. The nominal 95% interval is [+57.0, +62.8]. The study is archived with a DOI, and it also publishes the method, analysis/reproduction scripts, and verification procedures.
Speculative decoding is a generation method in which a draft proposes the next token and the main model verifies it. The repository also records results for draft-dflash (PR #27342), noting that throughput was 51.9% higher but that the results were not separated from draft-mtp. Decoding energy was reported as 37.1% lower, but the measurements were not calibrated with an external meter. The speed and energy figures are therefore server-reported results from this experiment and do not guarantee performance in general environments.
The inputs are 25 prompts composed for the purpose, so they do not represent the full set of actual requests. Under a 1600-token limit, the outputs for 23~25 of the 25 inputs differed from the default serial greedy decoding. This result shows that output equivalence cannot be judged by speed gains alone. The pre-registration and reproduction materials help with review, but performance claims must be interpreted within the limited inputs and the uncalibrated energy measurements.