الملخص
Speculative decoding accelerates autoregressive generation by using a lightweight draft to propose multiple tokens for parallel verification. However, existing methods often require an additional draft model or weight representation, introducing non-negligible memory overhead on resource-constrained devices. Self-speculative approaches reduce this overhead, yet still face trade-offs between draft quality, target quality, and storage efficiency. We propose BitNest, a bit-nested speculative decoding framework that embeds a low-precision draft directly into the higher-precision target representation. Instead of deriving a draft from a predefined target, BitNest first constructs a strong low-precision base and then recovers the higher-precision target through residual refinement, enabling both models to share a single physical weight representation. BitNest further extends this progressive-precision design to the KV cache for long-context inference. Across multiple 7B--8B edge-friendly LLMs and diverse workloads, BitNest achieves an average speculative acceptance rate of 95.2% while closely preserving higher-precision model quality, and delivers 1.48--1.61x end-to-end speedup over FP16 autoregressive decoding. On the LLaMA models supported by all representative self-speculative baselines, BitNest also achieves consistently competitive or higher decoding speedup.
الكلمات المفتاحية
الموضوع
بيانات النشر
- المجلة
- غير متاح
- وصول مفتوح
- وصول مفتوح أخضر
اقتبس هذه المقالة
APA 7
Yang, C., Cheng, N., Akbari, A., Tan, Q., Zhu, Q., Zhang, C., Yang, C., Wang, Y., Niu, W., Wang, J., Lu, J., & Yuan, G. (2026). BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration. https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration
MLA 9
Yang, Chence, et al. "BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration." https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration.
شيكاغو (المؤلف–التاريخ)
Yang, Chence, Ningxi Cheng, Arash Akbari, Qitao Tan, Qingchan Zhu, Ci Zhang, Changdi Yang, Yanzhi Wang, Wei Niu, Jinhui Wang, Jin Lu, and Geng Yuan. 2026. "BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration." https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration.
هارفارد
Yang, C., Cheng, N., Akbari, A., Tan, Q., Zhu, Q., Zhang, C., Yang, C., Wang, Y., Niu, W., Wang, J., Lu, J. and Yuan, G. (2026) 'BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration', Available at: https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration.
فانكوفر
Yang C, Cheng N, Akbari A, Tan Q, Zhu Q, Zhang C, et al. BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration. https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration
IEEE
C. Yang, N. Cheng, A. Akbari, Q. Tan, Q. Zhu, C. Zhang, C. Yang, Y. Wang, W. Niu, J. Wang, J. Lu, and G. Yuan, "BitNest: Bit-Nested Speculative Decoding for Memory-Efficient LLM Inference Acceleration," https://omanscience.com/ar/articles/bitnest-bit-nested-speculative-decoding-for-memory-efficient-llm-inference-acceleration.