Skip to content
← Quantization & Numerics / 08
hardNewNVIDIAMetaQualcomm

Which layers should stay in higher precision after quantization, and how do you find them with a sensitivity sweep?

A few matrices cause most of the damage when a model is quantized, and keeping them wide costs little memory. The answer builds the sweep that finds them, prices it in GPU hours, and explains why the sum of single-layer damages is not the damage of the mix.

Updated Oct 2026 · Worked answers on inference mechanisms, with calculations, implementation details, and follow-up questions.

A few matrices cause most of the damage when a model is quantized, and keeping them wide costs little memory. The answer builds the sweep that finds them, prices it in GPU hours, and explains why the sum of single-layer damages is not the damage of the mix.

an account raises the free count per topic
UP NEXT ON YOUR JOURNEY
Next in this trackIn a block-scaled FP8 GEMM like DeepSeek V3's, where are the scales applied, and why must accumulation be promoted to FP32?Next in this trackGGUF k-quants and EXL2 mix bit-widths inside one model. What do they trade, and where do they fit next to datacenter GPU formats?Next in this trackThe same prompt gives different logits at batch 1 and batch 64 on an FP8 server. Why, and how can the kernels prevent it?
DISCUSSION · 0

Nobody has commented on this answer. Add a correction or another way to work it.