qwen3.6-speculative-decoding-rtx3090
A reproducibility study measuring llama.cpp speculative decoding for a quantized Qwen3.6-35B model on a single RTX 3090, with re-derived figures, a coverage probe and a public errata of the study's own retracted claims.
This repository re-measures llama.cpp speculative decoding for the Qwen3.6-35B-A3B UD-Q4_K_XL GGUF model on one RTX 3090, split into two tiers that the authors say must not be read as one body of evidence: an archival tier of single-request decode microbenchmarks from three earlier llama.cpp commits, and a controlled tier of over 3,000 arm-runs collected across 74 directories against a specific post-merge commit, each with a matched no-speculation baseline, verified thinking suppression, verified concurrent-request timing, full per-request text and token ids, and continuous GPU telemetry.
The published figures are re-derived from the committed raw data by a checker that fails on drift, and a coverage probe reports how many of the published figures it actually covers, so a reader can tell which numbers are checked automatically. The project documents concrete limits rather than a single headline number: the same configuration measured twelve times in one day spans a 9.4 percentage-point range on byte-identical output, and every thinking-off comparison that let each arm stop when it liked is confounded by output length; a public errata file lists the study's own retracted claims from its audit.