T2Quant allocates precision by task and timestep in diffusion LMs
Original title: Different Tasks, Different Timesteps: Temporal Task-Aware Quantization for Diffusion Language Models
On three diffusion language models and six benchmarks, the method raises average accuracy by up to 6.7 points at a fixed bit budget.
The paper argues that quantization sensitivity in diffusion language models depends on both the downstream task and the denoising timestep. It says those patterns can be predicted without ground-truth labels or task-specific metrics. T2Quant uses a label-free sensitivity signal to assign activation precision across denoising steps under a fixed bit-width budget, and the authors report consistent gains over existing quantization methods.
- Experiments cover three diffusion language models and six benchmarks.
- Reported accuracy gains reach 6.7 percentage points under the same bit-width budget.
- Precision is allocated to activations across denoising steps, not a single static setting.
From the original article
1 min read at huggingface.co
Diffusion Language Models (DLMs) introduce unique quantization challenges due to their iterative denoising process. In this work, we reveal that quantization sensitivity is both taskand time-dependent, with different downstream tasks exhibiting distinct precision requirements across denoising steps. We further show that these patterns can be predicted without groundtruth labels or task-specific evaluation metrics.
By Ofek Avivi, Eliron Rahimi, Avi Mendelson · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.