AI BriefingMBD-LM: Optimizing Parallel Generation for Diffusion-Based Language Models
Key point
The paper proposes MBD-LM, a Multi-Block approach designed to improve the parallel decoding performance of Block Diffusion Language Models (BD-LMs).
Details
Existing Block Diffusion Language Models (BD-LMs) have limitations in parallel processing and variable-length generation due to their single-block-based generation method. To address this, the researchers propose Multi-Block Diffusion Language Models (MBD-LMs).
The key features of MBD-LMs are as follows:
- Multi-block Teacher Forcing (MultiTF): Combines the advantages of existing Teacher Forcing and Diffusion Forcing, reflecting the non-uniform noise patterns that occur during inference in the training process.
- Block Buffer mechanism: Fixes the input shape while maintaining Prefix-cache reuse, translating decoding parallelism into actual Wall-clock acceleration.
Experimental results show that the MBD-LLaDA2-Mini model significantly improved the average Tokens Per Forward pass (TPF) from 3.47 to 6.19, and accuracy also improved from 79.95% to 81.03%. In particular, when combined with DMax, it showed high efficiency with a TPF reaching 9.34.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Underdog Releases 'Husky', a Model-Specific Inference Engine Up to 4.5x Faster than MLX