QLoRA · AI.WIKI

01 · 两个省资源旋钮不能绑在一起01 · Two resource-saving dials should not be glued together

把基座从 16 位压到 4 位，为什么不会自动把 LoRA 的 r 也变小？反过来调 r，又为什么不改变基座文件精度？Why does compressing the base from 16-bit to 4-bit not automatically reduce LoRA rank—and why does changing rank not alter base precision?

02用位宽与 rank r 两个独立控件驱动显存、参数量与误差Use Independent Bit-width and Rank Controls to Drive Memory, Parameters, and Error

位宽 ⟂ 低秩precision ⟂ low rank

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、得分与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。All probabilities, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

QLoRA

QBASE

量化冻结基座Quantized frozen base

位宽减少冻结权重存储，不直接训练基座。Bit width reduces frozen-weight storage without training the base.

LORA

LoRA 增量LoRA delta

A、B 矩阵保持可训练。The A and B matrices remain trainable.

BITS

位宽杠杆Bit-width lever

16/8/4 位主要改变存储与量化误差。16/8/4 bits mainly affect storage and quantization error.

RANK

秩杠杆Rank lever

r 主要改变增量参数与近似容量。r mainly affects delta parameters and approximation capacity.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

output = dequantize(W_q)x + B_rA_rx

点击位宽只让灰色基座层变深；拖 r 只让红色适配器变宽，两个指标不会偷偷联动。Bit-width buttons change only the gray base layers; moving r changes only the red adapters. The controls never silently move each other.

压缩基座Quantize base减少冻结权重存储reduce frozen-weight storage

训练增量Train deltaLoRA A、B 保持高精度keep LoRA A and B trainable

组合推理Compose inference两类误差共同验收evaluate both error sources

05同时选择 4-bit 基座与 r = 1Combine a 4-bit Base with r = 1

主动制造失败Create a failure

边界实验Boundary experiment

位宽和秩都压到最低时资源最省，但量化误差与低秩近似误差会叠加。Minimizing both precision and rank saves the most resources, but quantization and low-rank approximation errors accumulate.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ QLoRA 就是 4-bit LoRA，两个旋钮永远绑定。QLoRA is 4-bit LoRA with permanently linked dials.→ ✓ 位宽与 r 是独立设计与验收变量。Bit width and r are independent design and evaluation variables.

✗ 量化基座也会像 LoRA 一样被训练。The quantized base is trained like LoRA.→ ✓ QLoRA 路线中基座冻结，训练的是低秩增量。In QLoRA, the base stays frozen while the low-rank delta trains.

✗ 显存最低的配置就是最佳配置。The lowest-memory configuration is always best.→ ✓ 还要同时验收任务质量与两类误差。Task quality and both error sources still need evaluation.

06 · 一句话带走06 · One line to keep

QLoRA 用量化位宽压缩冻结基座，用 LoRA 秩控制可训练增量；两个杠杆独立，误差却会在最终输出中合流。QLoRA compresses the frozen base with bit width and controls the trainable delta with LoRA rank. The levers are independent, but their errors meet in the final output.