Quantization Without Degradation: Preserving Multi-Step Reasoning at INT4 & FP8
KEY TAKEAWAYS
- Standard post-training quantization often destroys fragile activation outliers in deep reasoning attention heads.
- Activation-aware calibration with Solace 1.0 traces retains 98.6% of FP16 accuracy on Math-500 in INT4 AWQ.
- GGUF and FP8 checkpoints are now available on Hugging Face.
Quantizing reasoning models has notoriously high loss margins compared to conversational chatbots. Chain-of-thought tokens require high dynamic range to avoid error cascades.
In this technical note, we detail our activation calibration pipeline for producing zero-degradation FP8 and INT4 AWQ checkpoints.
FP16 Baseline Accuracy (Math-500): 84.2%
Standard RTN INT4: 61.8% (-22.4% drop)
Solace Calibrated AWQ INT4: 83.1% (-1.1% delta)
Check out the full quantization recipes in our Docs Guide.