DistillSpec: Improving Speculative Decoding via Knowledge Distillation
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:UVGNT6JVrecord.jsonopen to challenge →
read the original abstract
Speculative decoding (SD) accelerates large language model inference by employing a faster draft model for generating multiple tokens, which are then verified in parallel by the larger target model, resulting in the text generated according to the target model distribution. However, identifying a compact draft model that is well-aligned with the target model is challenging. To tackle this issue, we propose DistillSpec that uses knowledge distillation to better align the draft model with the target model, before applying SD. DistillSpec makes two key design choices, which we demonstrate via systematic study to be crucial to improving the draft and target alignment: utilizing on-policy data generation from the draft model, and tailoring the divergence function to the task and decoding strategy. Notably, DistillSpec yields impressive 10 - 45% speedups over standard SD on a range of standard benchmarks, using both greedy and non-greedy sampling. Furthermore, we combine DistillSpec with lossy SD to achieve fine-grained control over the latency vs. task performance trade-off. Finally, in practical scenarios with models of varying sizes, first using distillation to boost the performance of the target model and then applying DistillSpec to train a well-aligned draft model can reduce decoding latency by 6-10x with minimal performance drop, compared to standard decoding without distillation.
This paper has not been read by Pith yet.
Forward citations
Cited by 16 Pith papers
-
Weak-to-Strong Generalization via Direct On-Policy Distillation
Transferring the log-ratio of a small model's pre-RL and post-RL checkpoints provides a dense implicit reward that improves stronger student models at a fraction of the cost of direct RL.
-
SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting
SpecBlock achieves 8-13% higher mean speedup than EAGLE-3 at 44-52% drafting cost via block-iterative drafting with hidden-state inheritance, dynamic rank-head branching, valid-prefix masking, and optional cost-aware ...
-
Knowledge Distillation for Visual Autoregressive Models
VarKD is a distillation framework for visual AR models that uses student samples and selective teacher supervision to reduce token ambiguity, outperforming prior baselines on ImageNet.
-
DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding
DFlare replaces DFlash's shared fused representation with per-draft-layer attention to distinct target-layer combinations, enabling deeper drafts and 2.4M training samples for 5-11% higher speedups than DFlash on Qwen...
-
Draft-OPD: On-Policy Distillation for Speculative Draft Models
Draft-OPD applies on-policy distillation via target-assisted generation and error replay to train speculative draft models, yielding over 5x lossless acceleration and gains over EAGLE-3 and DFlash.
-
Performance-Driven Policy Optimization for Speculative Decoding with Adaptive Windowing
PPOW uses window-level RL with cost-aware speedup and proximity rewards plus adaptive divergence-aware windowing to reach 6.29-6.52 acceptance lengths and 3.39-4.36x speedups in speculative decoding.
-
PARD-2: Target-Aligned Parallel Draft Model for Dual-Mode Speculative Decoding
PARD-2 uses Confidence-Adaptive Token optimization to align draft model training with acceptance length in speculative decoding, enabling dual-mode operation and up to 6.94x lossless speedup on Llama3.1-8B.
-
SpecBlock: Block-Iterative Speculative Decoding with Dynamic Tree Drafting
SpecBlock achieves 8-19% higher speedup than EAGLE-3 in LLM speculative decoding by using repeated block expansions with hidden-state inheritance, a dynamic rank head, and a valid-prefix training mask.
-
Position-Aware Drafting for Inference Acceleration in LLM-Based Generative List-Wise Recommendation
PAD-Rec augments standard draft models with item-position and step-position embeddings plus learnable gates, delivering up to 3.1x wall-clock speedup and 5% average gain over strong speculative-decoding baselines on f...
-
Multi-Token Prediction via Self-Distillation
Self-distillation turns pretrained autoregressive LMs into multi-token predictors that decode over 3x faster with under 5% accuracy drop on GSM8K.
-
EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty
EAGLE resolves feature-level uncertainty in speculative sampling via one-step token advancement, delivering 2.7x-3.5x speedup on LLaMA2-Chat 70B and doubled throughput across multiple model families and tasks.
-
Acceptance Dynamics Across Cognitive Domains in Speculative Decoding
Empirical measurements across four NLP domains show task type is a stronger predictor of speculative decoding acceptance than tree depth, with chat uniquely achieving expected accepted length over 1 token per step.
-
ECHO: Elastic Speculative Decoding with Sparse Gating for High-Concurrency Scenarios
ECHO uses sparse gating and elastic budget pivoting in a super-tree structure to achieve up to 5.35x speedup for LLM inference under high concurrency.
-
A Survey of On-Policy Distillation for Large Language Models
A survey that formalizes on-policy distillation as f-divergence minimization over student-sampled trajectories and organizes the literature along three design axes while linking it to KL-constrained RL.
-
A Survey on Efficient Inference for Large Language Models
The paper surveys techniques to speed up and reduce the resource needs of LLM inference, organized by data-level, model-level, and system-level changes, with comparative experiments on representative methods.
-
A Survey of On-Policy Distillation for Large Language Models
On-policy distillation reframes LLM knowledge transfer as iterative correction on student trajectories rather than single-pass imitation, with the survey organizing the field along divergence design, feedback sources,...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.