Quantização

A neural network uses numerical values for weights and intermediate activations. Quantization changes the representation of selected values, often reducing the number of bits used for storage. A weight-only scheme should be distinguished from one that also quantizes activations. Bit counts provide an ideal payload calculation: storing the same number of weights at four bits instead of sixteen uses one-quarter of the raw packed-weight bits. That arithmetic does not include scale metadata, unquantized values, activations, caches, or temporary workspace, so it is not a promise about total device memory. Quantization can introduce rounding or other approximation error. Its effect on predictions needs evaluation rather than an assumption that every compressed model stays identical to its baseline. Calibration procedures use representative inputs to guide decisions. GPTQ uses approximate second-order information for post-training weight quantization. AWQ uses activation observations to identify salient weight scaling choices; activation-aware describes information used for weight quantization, not a rule that all activations are themselves stored as low-bit integers. Lower-bit packing can reduce transferred bytes, but runtime gains also depend on kernels, hardware, and conversion overhead. Storage and arithmetic precision can differ: QLoRA, for example, distinguishes low-bit stored weights from higher-precision computation. An affine integer representation offers a useful illustration. With a stated reconstruction rule of scale times code minus zero point, code six, scale one-half, and zero point two reconstruct the value two. Quantization-aware training instead exposes training to the effects of the chosen quantization procedure so learned values can adapt; post-training methods act on an already trained starting model.