thc1006/qwen3.8-speculative-decoding-rtx3090

thc1006/qwen3.8-speculative-decoding-rtx3090 是一个研究仓库，在单张 RTX 3090 上对 Qwen3.8-27B 的推测解码性能与输出差异进行了预先注册比较。

GitHub 项目说明: Preregistered, DOI-archived llama.cpp study: Qwen3.8-27B on one RTX 3090 (sm_86). draft-mtp at --spec-draft-n-max 2 is +59.8% [+57.0, +62.8] server-reported decode over 25 purposive prompts; draft-dflash (PR #27342) +51.9%, not separated from it. Decode energy -37.1%, uncalibrated. 23-25 of 25 prompts diverge from serial greedy at 1600 tokens.

星标

★ 3

复刻

0

语言

Python

许可证

MIT

最新版本

v1.0.4