Post-training quantization: GPTQ, AWQ, FP8
Trained weights are too fat to serve. Post-training quantization compresses them after training, using calibration data to choose rounding that preserves the layer's function: GPTQ via iterative second-order error compensation, AWQ via activation-aware scaling, FP8 via calibrated ranges.
GPTQ
Quantize columns one at a time, redistributing each column's rounding error into the not-yet-quantized weights via the inverse Hessian. 4-bit with group size 128 is the sweet spot.
AWQ
Protect the ~1% of salient weight channels (large activations) by per-channel scaling instead of mixed precision — hardware-friendly, no outliers stored separately.
FP8 Calibration
Dynamic ranges need a calibration pass; too few calibration samples inflates KL divergence against the bf16 reference.
4-bit, group 128: err ≈ 0.0028 → ~98.9% quality retained. A 70B model: 140 GB at bf16 vs 35 GB at 4-bit — the difference between a node and a laptop.
- Quantize the final aligned model, not an earlier checkpoint — alignment is where behavior lives, and PTQ after alignment preserves it.
- Always eval the quantized model on your task suite: per-layer error compounds non-uniformly, and 4-bit can cost 1–3 points on reasoning-heavy sets.