{"id":8202,"date":"2025-02-20T01:00:00","date_gmt":"2025-02-20T01:00:00","guid":{"rendered":"https:\/\/alphacrc.com\/?post_type=insight&p=8202"},"modified":"2025-04-02T10:06:56","modified_gmt":"2025-04-02T10:06:56","slug":"ai-localization-quantization","status":"publish","type":"insight","link":"https:\/\/alphacrc.com\/fr\/insight\/ai-localization-quantization\/","title":{"rendered":"AI localization: Using quantization to improve adaptability"},"content":{"rendered":"\n
With artificial intelligence tools continuing to make inroads into the localization industry, linguistic service providers (LSPs) need to re-evaluate integration methods to reduce turnaround times while preserving translation quality.<\/p>\n\n\n\n
Implementing AI localization effectively can be challenging, especially when aiming to do more complex tasks than simply getting an LLM-powered translation. There are often restrictions from popular LLMs sourced through APIs (such as OpenAI) and running larger and more capable open-source models on-site can be expensive.<\/p>\n\n\n\n
AI localization and the open-source problem<\/h2>\n\n\n\n
Open-source models can offer businesses a much greater degree of customization and the ability to avoid iterative costs, but they require high-grade hardware in order to run, specifically a good graphics card (GPU) to run calculations, though it is possible, albeit slower, to use the central processing unit (CPU) for this.<\/p>\n\n\n\n
Running Llama 3 at full precision means running it at float 32 (FP32), a computer number format that typically occupies 32 bits (four bytes) of memory. The lightest Llama 3 model contains 8 billion parameters, each operating at FP32 precision. Thus:<\/p>\n\n\n\n
8,000,000,000 parameters x 4 bytes (memory usage per parameter at FP32) = 32,000,000,000 byte (roughly 30GB) graphics card memory requirement.<\/p>\n\n\n\n
But it would be prohibitively expensive to run this Llama 3 on internal hardware. If businesses want to run open-source models such as Llama to perform more varied and nuanced AI experiments, they need to look at quantization techniques.<\/p>\n\n\n
\n<\/figure><\/div>\n\n\n
What is quantization?<\/h2>\n\n\n\n
In this context, quantization refers to the process of reducing the precision of a model\u2019s weights, which compress the model by reducing the number of bits required to represent each parameter.<\/p>\n\n\n\n
Quantization aims to reduce the memory footprint required to run resource-intensive open-source LLMs, even enabling use of edge devices such as phones. Reducing precision also results in faster computation and lower energy consumption.<\/p>\n\n\n\n
There are a range of potential quantization methods suitable for AI localization, including post-training quantization, quantization-aware training and dynamic quantization, with the end effect meaning lower requirements to run open-source models.<\/p>\n\n\n\n
Going back to Llama 3, quantization offers a reduction in precision from FP32 to FP16 or even float point 8 (FP8). Running Llama 3\u2019s lightest model at FP16 would mean:<\/p>\n\n\n\n
8,000,000,000 parameters x 2 bytes (memory usage per parameter at FP16) = 16,000,000,000 byte (roughly 15GB) graphics card memory requirement.<\/p>\n\n\n\n
Significant gains such as this can result in massive savings for LSPs looking to experiment with AI integration more freely than is possible through APIs from companies such as OpenAI.<\/p>\n\n\n\n
What are the trade-offs?<\/h2>\n\n\n\n
\n
\n
To start with, FP16 has a smaller dynamic range and lower precision compared to FP32, which can result in numerical instability. Gradients could also become less accurate, opening up the potential for slower convergence. This would, of course, be more of an issue during training than inference.<\/p>\n\n\n\n
That said, the practical results of this shouldn\u2019t be underestimated. In a translation environment, this could mean subtle nuances or infrequent word associations are lost, or output could become increasingly inaccurate.<\/p>\n\n\n\n
Let\u2019s take a look at some potential outcomes of quantization at each step of the LLM-based translation process.<\/p>\n<\/div>\n\n\n\n
\n<\/figure>\n<\/div>\n<\/div>\n\n\n\n
LLM translation and potential quantization issues<\/h2>\n\n\n\n
There are four key stages to the LLM-based translation process, each of which could experience issues stemming from quantization if not appropriately addressed.<\/p>\n\n\n\n
Input processing<\/h3>\n\n\n\n
\n
Tokenization: Input text is broken down into smaller units called tokens.<\/li>\n\n\n\n
Encoding: Tokens are converted into numerical representations that the model can process.<\/li>\n<\/ul>\n\n\n\n
Potential issues<\/p>\n\n\n\n
\n
Encoding: Quantization can lead to errors in the embedding layer, where continuous space representations of tokens are approximated. This could potentially result in less accurate embeddings, which would limit the model\u2019s understanding of the input text.<\/li>\n<\/ul>\n\n\n\n
Contextual understanding<\/h3>\n\n\n\n
\n
Contextual embedding: The model captures the meaning of each token in relation to those surrounding it.<\/li>\n\n\n\n
Attention mechanisms: This ensures important information is retained over irrelevant content.<\/li>\n<\/ul>\n\n\n\n
Potential issues<\/p>\n\n\n\n
\n
Contextual embedding: Quantization can lower contextual precision, affecting the model\u2019s ability to capture nuances and relationships between tokens.<\/li>\n\n\n\n
Attention mechanisms: The model may focus on less relevant parts of the text.<\/li>\n<\/ul>\n\n\n\n
Translation generation<\/h3>\n\n\n\n
\n
Decoding: Translation is generated one token at a time.<\/li>\n\n\n\n
Beam search: This keeps track of multiple possible translations and selects the best one based on a scoring mechanism.<\/li>\n<\/ul>\n\n\n\n
Potential issues<\/p>\n\n\n\n
\n
Decoding: Precision of generated tokens may be lower. The model may therefore generate tokens that are close but not exactly correct.<\/li>\n\n\n\n
Beam search: The scoring mechanism may be affected, leading to suboptimal translation path selection.<\/li>\n<\/ul>\n\n\n\n
Output processing<\/h3>\n\n\n\n
\n
Detokenization: Generated tokens are converted back into human-readable text.<\/li>\n\n\n\n
Post-processing: Additional steps may be taken to improve the translation, such as correcting grammar.<\/li>\n<\/ul>\n\n\n\n
Potential issues<\/p>\n\n\n\n
\n
Post-processing: As a step in itself, post-processing will be less affected by quantization, but any errors introduced in previous stages can propagate and would become more noticeable here.<\/li>\n<\/ul>\n\n\n\n
\n
\n<\/figure>\n<\/div>\n<\/div>\n\n\n\n
Mitigating the effects of quantization<\/h2>\n\n\n\n
For AI localization, it\u2019s vital that accuracy and context need to be preserved, meaning that the potential impacts of quantization must be mitigated. There are several methods that could aid with this.<\/p>\n\n\n\n
Mixed-precision quantization<\/h3>\n\n\n\n
While quantization can significantly reduce hardware requirements, it doesn\u2019t need to be applied across the entire model. Mixed-precision quantization keeps critical layers at higher prevision levels such as FP32, while reducing other layers to lower points such as FP16.<\/p>\n\n\n\n
Quantization-aware training<\/h3>\n\n\n\n
Businesses can also consider training the model with quantization in mind. By simulating the effects of quantization during training, the model can learn how best to compensate for the lower precision levels. This typically results in better performance when compared to processes such as raw post-training quantization.<\/p>\n\n\n\n
Post-quantization fine-tuning<\/h3>\n\n\n\n
By fine-tuning on a relevant dataset after the quantization process is complete, the model can better adjust to its lower precision and recover some of its lost performance.<\/p>\n\n\n\n
Error-feedback mechanisms<\/h3>\n\n\n\n
Implementing error-feedback mechanisms, such as the employment of human specialists can help the model to correct quantization errors during the inference stage of translation. The model can maintain the quality of translated or localized content from feedback generated over time.<\/p>\n\n\n\n
![\"Group](\"https:\/\/alphacrc.com\/wp-content\/uploads\/2025\/02\/AI-dev-1024x622.png\")
<\/figure><\/div>\n\n\n![\"Creating](\"https:\/\/alphacrc.com\/wp-content\/uploads\/2025\/02\/Gears-1024x1001.png\")
<\/figure>\n<\/div>\n<\/div>\n\n\n\n![\"Working](\"https:\/\/alphacrc.com\/wp-content\/uploads\/2025\/02\/Puzzle-dance-1024x576.png\")
<\/figure>\n<\/div>\n<\/div>\n\n\n\n