Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2023
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866916603725086720 |
|---|---|
| author | Zhang, Jun Wang, Jue Li, Huan Shou, Lidan Chen, Ke Chen, Gang Mehrotra, Sharad |
| author_facet | Zhang, Jun Wang, Jue Li, Huan Shou, Lidan Chen, Ke Chen, Gang Mehrotra, Sharad |
| contents | We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting. Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM. Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99$\times$. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2309_08168 |
| institution | arXiv |
| publishDate | 2023 |
| record_format | arxiv |
| spellingShingle | Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding Zhang, Jun Wang, Jue Li, Huan Shou, Lidan Chen, Ke Chen, Gang Mehrotra, Sharad Computation and Language We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting. Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM. Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99$\times$. |
| title | Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2309.08168 |