Why it matters
AWQ is often the top choice for INT4 quantization now. Understanding it complements GPTQ knowledge.
Advertisement
The architecture
Analyze activations: find channels with large activation magnitudes.
Protect corresponding weights: scale them up before quantization so quantization error is smaller relative to their value.
Advertisement
How it works end to end
Salient weights: those in channels with large activations dominate output. Preserving them preserves quality.
Grouped scales: still uses per-group scales for storage efficiency.
No optimization: unlike GPTQ, doesn't solve per-layer optimization. Simpler.
Quality: often matches or beats GPTQ for LLMs.