Functional Reconstruction Improves MLA Draft Models for Speculative Decoding

Weiye Shi, Fanxu Meng, Muhan Zhang· July 31, 2026 View original

Key takeaways

  • Direct MHA/GQA-to-MLA conversion can degrade speculative decoding performance due to attention-function errors.
  • Functional Reconstruction optimizes MLA draft models to reproduce original attention behavior, improving draft token acceptance.
  • This method is converter-agnostic and does not require verifier supervision.
  • It enables more efficient long-context LLM inference by combining MLA's memory benefits with speculative decoding's speed.

Who benefits

AI/ML DevelopmentCloud ComputingSoftware EngineeringData ScienceResearch & Academia

Summary

Researchers propose Functional Reconstruction, an end-to-end method to optimize Multi-head Latent Attention (MLA) draft models for speculative decoding. This approach addresses the performance degradation caused by direct conversion from Multi-head Attention (MHA) or Grouped-Query Attention (GQA) to MLA, significantly improving draft token acceptance rates.

Multi-head Latent Attention (MLA) is gaining traction for long-context Large Language Model (LLM) inference due to its efficiency in reducing memory traffic by using compact latent states instead of a growing Key-Value (KV) cache. However, many powerful open-source LLMs utilize Multi-head Attention (MHA) or Grouped-Query Attention (GQA), requiring conversion to MLA for these efficiency benefits. This conversion, particularly when combined with speculative decoding, can lead to a sharp reduction in draft token acceptance, undermining the acceleration benefits. The issue stems from attention-function errors introduced during direct MHA/GQA-to-MLA conversion, which, while potentially tolerable for standalone generation, significantly impact the agreement between draft proposals and target verification in speculative decoding. To counter this, a new method called Functional Reconstruction is introduced. This end-to-end procedure optimizes each converted MLA attention module to precisely reproduce the post-output-projection response of its original MHA/GQA counterpart on calibration hidden states. Functional Reconstruction is converter-agnostic and does not require verifier logits or supervision, preserving the converted cache and inference graph. Extensive evaluations across various model families, converters, backends, and tasks demonstrate that this method materially improves draft token acceptance in a significant majority of cases, making speculative decoding with MLA more effective.

Why it matters

For professionals working with LLMs, especially those focused on long-context inference and efficiency, improving speculative decoding performance is crucial. This research offers a way to leverage MLA's memory benefits without sacrificing the speedups from speculative decoding, leading to faster and more cost-effective LLM deployments.

How to implement this in your domain

  1. 1Evaluate current MHA/GQA-to-MLA conversion pipelines for LLMs, especially when used with speculative decoding, to identify potential performance bottlenecks.
  2. 2Implement the Functional Reconstruction method to optimize converted MLA attention modules, aiming to reproduce the original MHA/GQA attention function.
  3. 3Test the impact of Functional Reconstruction on draft token acceptance rates and overall speculative decoding speedup for long-context LLM tasks.
  4. 4Integrate this post-conversion optimization step into LLM deployment workflows to enhance inference efficiency.

Original post by Weiye Shi, Fanxu Meng, Muhan Zhang

"arXiv:2607.27269v1 Announce Type: new Abstract: Multi-head latent attention (MLA) is increasingly important for long-context LLM inference because compact latent states replace the growing key-value (KV) cache and reduce decoding memory traffic. Yet most capable open checkpoints…"

View on X

Originally posted by Weiye Shi, Fanxu Meng, Muhan Zhang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses