TR-PTQ: High-Accuracy Integer-Only Transformer Post Training Quantization via Taylor Region Reformulation
Post-training quantization (PTQ) enables efficient deployment, yet transformer architectures remain challenging to quantize due to nonlinear layers.
Key points
- While existing methods attribute accuracy loss to insufficient numerical precision, often necessitating floating-point fallbacks, we demonstrate that degradation is actually driven by specific structural error sources.
- We find that learned scale parameters in normalization layers and compounded approximations in GELU are the primary error contributors, whereas SoftMax remains inherently robust to aggressive quantization.
- To address these bottlenecks, we introduce TR-PTQ, a unified integer-only formulation using shared Taylor Region (TR) exponential and logarithm primitives.
- Combined with a calibration-free, outlier-aware optimization for LayerNorm parameters, our method eliminates the need for floating-point hardware units for nonlinearities, achieving less than 1.5% absolute accuracy degradation across vision and language benchmarks.
Sources (1)
- [1]TR-PTQ: High-Accuracy Integer-Only Transformer Post Training Quantization via Taylor Region ReformulationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 12:39 PM
Post-training quantization (PTQ) enables efficient deployment, yet transformer architectures remain challenging to quantize due to nonlinear layers.
While existing methods attribute accuracy loss to insufficient numerical precision, often necessitating floating-point fallbacks, we demonstrate that degradation is actually driven by specific structural error sources.
Extractive summary: sentences quoted from the sources.