Self-Speculative Decoding (Zhang 2023年)（セルフスペキュレイティブデコーディング）

概要

Self-Speculative Decoding は、2023年Zhang et al. (HKU+Microsoft)発表Self-Speculative Decoding・Industry-leading self-speculative decoding LLM 2023年 + Industry-leading 1.5-1.99× inference speedup + Industry-leading no draft model position確立。Self-Speculative Decoding specifications = Industry-leading self-speculative decoding LLM (Industry-leading Self-Speculative 2023 + Industry-leading self-speculative decoding LLM + Industry-leading Self-Speculative flagship) + Industry-leading same model skip layers as draft (Industry-leading same model skip selected intermediate layers as draft) + Industry-leading 1.5-1.99× inference speedup (Industry-leading 1.5-1.99× inference speedup vs vanilla) + Industry-leading no draft model required (Industry-leading no separate draft model required)。

主な特徴・仕組み

• Authors: Industry-leading Jun Zhang + Jue Wang + Huan Li + Lidan Shou + Ke Chen + Gang Chen + Sharad Mehrotra HKU + Microsoft + Zhejiang + UC Irvine
• Year: 2023年 (arXiv 2023年9月発表 + ACL 2024 accepted)
• Paper: Industry-leading "Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding" arXiv 2309.08168
• Method Type: Industry-leading self-speculative decoding draft+verify same model
• Key Innovation: Industry-leading same model skip layers as draft (no separate draft model)
• Draft Mechanism: Industry-leading skip selected intermediate layers for draft generation
• Verification: Industry-leading full model verification of draft
• Speedup: Industry-leading 1.5-1.99× inference speedup vs vanilla autoregressive
• Lossless: Industry-leading lossless guarantee (same output as vanilla)
• No Draft Model: Industry-leading no separate draft model required
• No Training: Industry-leading no additional training required (plug-and-play)
• Open Source: Industry-leading Self-Speculative GitHub dilab-zju/self-speculative-decoding
• Industry-Leading: Industry-leading self-speculative decoding LLM first pioneer

スペック比較表

具体例・対応製品

• Self-Speculative Decoding (2023年, Zhang et al. HKU+Microsoft): Industry-leading self-speculative decoding
• Industry-leading same model skip layers as draft: Industry-leading same model skip layers
• Industry-leading skip selected intermediate layers for draft: Industry-leading skip intermediate layers
• Industry-leading full model verification of draft: Industry-leading full model verification
• Industry-leading no additional training required plug-and-play: Industry-leading plug-and-play
• 競合 Medusa + EAGLE + Lookahead + SpecExec: Industry-leading speculative decoding competitors

自作PCでの選び方・注意点

Self-Speculative Decoding は「Industry-leading self-speculative decoding LLM + Industry-leading same model skip layers as draft」「Industry-leading 1.5-1.99× inference speedup + Industry-leading no draft model required」用途のIndustry-leading self-speculative decoding 2023年Zhang et al. HKU+Microsoft発表product。Industry-leading same model skip (Industry-leading same model skip selected intermediate layers as draft + Industry-leading skip layers for draft generation + Industry-leading no separate draft model required + Industry-leading no additional training required plug-and-play) で Industry-leading same model skip layers + Industry-leading skip intermediate layers + Industry-leading no separate draft + Industry-leading plug-and-play。Industry-leading full model verification (Industry-leading full model verification of draft + Industry-leading lossless guarantee same output as vanilla + Industry-leading draft & verify signature) で Industry-leading full model verification + Industry-leading lossless guarantee + Industry-leading draft & verify。Industry-leading 1.5-1.99× speedup (Industry-leading 1.5-1.99× inference speedup vs vanilla autoregressive + Industry-leading Self-Speculative speedup advantage + Industry-leading no draft model pioneer) で Industry-leading 1.5-1.99× speedup + Industry-leading Self-Speculative advantage + Industry-leading no draft model pioneer。Industry-leading plug-and-play (Industry-leading no additional training required + Industry-leading plug-and-play deployment + Industry-leading Self-Speculative GitHub dilab-zju/self-speculative-decoding open source + Industry-leading self-speculative decoding LLM first pioneer) で Industry-leading no training + Industry-leading plug-and-play + Industry-leading dilab-zju open source + Industry-leading first pioneer。但しIndustry-leading Medusa + EAGLE + Lookahead + SpecExec competition (Industry-leading Medusa multiple heads 2-3× 2024 + EAGLE feature-level 2.5-3.5× 2024 + Lookahead n-gram 1.5-2× 2024 + SpecExec 18.7× consumer GPU 2024 vs Self-Speculative same model skip 1.99× 2023 trade-off) で Industry-leading 4-speculative decoding method competitors vs Self-Speculative + Industry-leading same model skip layers + no draft model + no training plug-and-play + lossless guarantee + first pioneer 2023 unique advantage adoption alignment必須。

関連用語との違い

• vs Medusa (Multiple Heads 2024): Self-Speculativeはsame model skip layers + 1.99×・Medusaは multiple decoding heads + 2-3×
• vs EAGLE (Feature-level 2024): Self-Speculativeはsame model + plug-and-play・EAGLEはfeature-level + tree draft + training required
• vs Lookahead (N-gram 2024): Self-Speculativeはskip layers + lossless・Lookaheadはn-gram pool + Jacobi parallel

よくある質問（FAQ）

Q1: Self-Speculative vs Medusa 違いは？ A: Self-Speculative Decoding (Industry-leading same model skip selected intermediate layers as draft + skip layers for draft generation + full model verification of draft + lossless guarantee same output as vanilla + no separate draft model required + no additional training required plug-and-play + 1.5-1.99× inference speedup + HKU+Microsoft 2023) vs Medusa (Industry-leading multiple lightweight decoding heads + 4-5 heads parallel prediction + tree-based attention candidate verification + no separate draft model required + 2-3× inference speedup + Princeton+Together AI 2024)・Industry-leading same model skip + lossless + plug-and-play + 1.99× = Self-Speculative + Industry-leading multiple heads + tree attention + no draft + 2-3× = Medusa preference judgment。

Q2: Industry-leading same model skip layers + plug-and-play value は？ A: Industry-leading same model skip + plug-and-play (Industry-leading same model skip selected intermediate layers as draft + Industry-leading skip layers for draft generation + Industry-leading no separate draft model required + Industry-leading no additional training required plug-and-play + Industry-leading lossless guarantee same output as vanilla)。

Q3: Industry-leading 1.5-1.99× speedup + lossless guarantee value は？ A: Industry-leading 1.5-1.99× + lossless (Industry-leading 1.5-1.99× inference speedup vs vanilla autoregressive + Industry-leading lossless guarantee same output as vanilla + Industry-leading full model verification of draft + Industry-leading Self-Speculative draft & verify signature)。

まとめ

Self-Speculative Decoding = 2023年Zhang et al. HKU+Microsoft発表のself-speculative decoding LLM。Industry-leading same model skip selected intermediate layers as draft + Industry-leading skip layers for draft generation + Industry-leading full model verification of draft + Industry-leading lossless guarantee same output as vanilla + Industry-leading no separate draft model required + Industry-leading no additional training required plug-and-play + Industry-leading 1.5-1.99× inference speedup vs vanilla autoregressive + Industry-leading Self-Speculative GitHub dilab-zju/self-speculative-decoding open source + Industry-leading self-speculative decoding LLM first pioneer 2023 position確立。