Post-Training Quantization (PTQ) Algorithms for LLMs

While Post-Training Quantization (PTQ) offers the significant advantage of quantizing models without expensive retraining, applying traditional PTQ techniques directly to Large Language Models (LLMs) often results in unacceptable accuracy degradation. As we discussed earlier, LLMs possess unique scale and architectural properties, making them particularly sensitive to the precision reduction inherent in quantization. Standard PTQ methods, which typically determine quantization parameters (scale $s$ and zero-point $z$) based on the simple range (min/max) of weights or activations, struggle to preserve the information encoded in the massive parameter spaces of LLMs, especially when targeting aggressive low-bit formats like INT4.

This limitation prompted the development of more sophisticated PTQ algorithms specifically designed for LLMs. These methods go further than simple range estimation and incorporate insights about the model's structure and data flow to minimize quantization error more intelligently. Two prominent examples that have gained significant traction are GPTQ and AWQ.

GPTQ: Layer-wise Quantization with Error Compensation

GPTQ (Generative Pre-trained Transformer Quantizer) tackles the accuracy challenge by employing a more careful, layer-by-layer quantization strategy combined with error compensation. Instead of quantizing all weights in a layer simultaneously based on global statistics, GPTQ processes weights sequentially, often in small blocks or columns.

The core idea is based on minimizing the reconstruction error for the layer's output. When a weight (or block of weights) is quantized, it introduces an error. GPTQ attempts to correct this error by making small adjustments to the remaining, not-yet-quantized weights within the same layer. This compensation mechanism aims to locally counteract the perturbation caused by quantization.

Mathematically, for a given layer weight matrix $W$ and input $X$, we want to find a quantized weight matrix $W_q$ such that the output $W_q X$ is as close as possible to the original output $W X$. GPTQ approaches this by solving an optimization problem for each layer:

$$\underset{W_q}{\text{argmin}} \| WX - W_q X \|_F^2 = \underset{W_q}{\text{argmin}} \| (W - W_q) X \|_F^2$$

Subject to the constraint that the elements of $W_q$ belong to the target low-bit representation. GPTQ uses an iterative, greedy approach. It quantizes weights one by one (or column by column) and updates the remaining weights using information derived from the inverse Hessian matrix $(X X^T)^{-1}$ (or approximations thereof) to minimize the squared error. This allows the remaining weights to adapt and compensate for the error introduced by the quantization of previous weights.

A comparison showing how GPTQ iteratively updates weights during quantization, unlike naive methods that quantize all at once.

Because it actively compensates for quantization errors during the process, GPTQ often achieves significantly better accuracy preservation than simpler methods, especially at 4-bit precision, albeit at the cost of a more computationally intensive quantization procedure.

AWQ: Activation-aware Weight Quantization

Activation-aware Weight Quantization (AWQ) takes a different approach, motivated by the observation that not all weights in an LLM are equally important. AWQ hypothesizes that weights connected to activations with larger magnitudes are more critical for the model's performance. Standard quantization methods treat all weights equally, which can disproportionately harm these salient weights.

AWQ proposes a simple yet effective solution: protect the important weights by scaling them. It doesn't directly modify the quantization process itself but introduces a per-channel scaling factor for the weights before quantization.

The process works as follows:

1. Analyze Activations: Feed calibration data through the model and observe the distribution of activation magnitudes entering each linear layer.
2. Identify Salient Channels: Determine channels where activations consistently have large magnitudes. These correspond to weight channels that are considered more important.
3. Determine Scaling Factors: Calculate per-channel scaling factors for the weight matrix. The goal is to find scales $s_c$ such that quantizing $W_c / s_c$ (where $W_c$ is the weight channel) minimizes quantization error, particularly for the salient channels identified in step 2. This effectively reduces the quantization step size for important weights relative to less important ones.
4. Quantize Scaled Weights: Apply standard PTQ (like symmetric per-channel quantization) to the scaled weights $W' = W / s$.
5. Inference: During inference, the scaling is reversed. The operation becomes $(W'_q \cdot s) X$, where $W'_q$ is the quantized version of the scaled weights. This scaling factor $s$ is absorbed into the layer's computation, often without adding significant overhead.

Flow of the Activation-aware Weight Quantization (AWQ) process, highlighting the activation analysis and weight scaling steps.

AWQ's primary advantage is its simplicity and speed compared to GPTQ, as it avoids complex iterative updates. It relies on the strong heuristic that activation magnitudes correlate well with weight importance. While perhaps not always reaching the absolute peak accuracy of GPTQ in every scenario, AWQ provides a compelling balance between quantization speed, ease of implementation, and resulting model accuracy, making it another popular choice for PTQ in LLMs.

Choosing Between GPTQ and AWQ

Both GPTQ and AWQ represent significant advancements over naive PTQ for LLMs.

• GPTQ often yields better accuracy, particularly for very low bit counts (e.g., 3-bit or 4-bit), due to its direct error minimization approach. However, the quantization process itself is considerably slower.
• AWQ offers a much faster quantization process and provides strong accuracy by protecting salient weights. It's simpler and often easier to implement and tune.

The choice between them might depend on the specific model, the target bit-width, the available computational resources for the quantization step, and the required level of accuracy preservation.

These advanced PTQ algorithms are essential tools for efficiently deploying LLMs. By enabling effective quantization down to low bit-widths with manageable accuracy loss, they significantly reduce the computational and memory requirements for inference. In the next chapter, we will look at the practical toolkits and libraries available for applying algorithms like GPTQ and AWQ to real-world LLMs.