量化、蒸馏、剪枝的区别Quantisation vs Distillation vs Pruning
微型指南 · 约 8 分钟Guide · about 8 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
三条路线都在「让模型变小」，但动的对象不同：量化降精度、蒸馏换学生、剪枝删结构。All three routes make a model smaller, but they touch different things: quantisation lowers precision, distillation swaps in a student, pruning removes structure.
量化：把权重从 FP16 降到 4/8 bit，模型结构不变；显存下降，质量取决于档位与校准数据。
蒸馏：用大模型（教师）的输出训练小模型（学生）；换的是模型，不是精度。
剪枝：删掉部分权重、通道或层；需要再训练/微调，结构真的变小。
代价不同：量化最省事但可能掉质量；蒸馏要训练数据与算力；剪枝要重训且可能破坏能力。
组合顺序：先蒸馏得到小模型，再量化部署；剪枝后通常要微调，再量化。
验收：同任务集对比质量、显存、延迟与成本；不要只比体积。
Quantisation: lower weights from FP16 to 4/8 bit while keeping the architecture; VRAM drops, quality depends on the tier and calibration data.
Distillation: train a small student on a large teacher’s outputs; you change the model, not the precision.
Pruning: remove weights, channels or layers; it needs retraining/fine-tuning and the structure really shrinks.
Costs differ: quantisation is easiest but may lose quality; distillation needs data and compute; pruning needs retraining and may damage capability.
Order: distil to a smaller model, then quantise for deployment; after pruning, fine-tune, then quantise.
Acceptance: compare quality, VRAM, latency and cost on the same task set; do not compare size alone.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
把同一模型权重换成低精度格式是量化；训练另一个学生模型是蒸馏；删除部分结构是剪枝。比较交付物和训练需求，不能因为文件都变小就认为三者可以互换。Lower-precision weights are quantization; training a student is distillation; removing structure is pruning. Compare artifacts and training needs rather than treating every smaller file as the same method.
- GPTQ 论文（arXiv 2210.17323）GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers ↗
- Hinton 等 · Distilling the Knowledge（arXiv 1503.02531）Hinton et al., Distilling the Knowledge in a Neural Network ↗
- Deep Compression（arXiv 1510.00149）Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding ↗
AI 常见误区总索引AI Misconceptions: Index · 量化会让模型变笨一半Quantisation Halves the Quality