Azerbaijani LLM Training Accelerates with Faster Throughput and Leaner Memory

In six weeks, Azerbaijani LLM training delivered a 23 percent throughput boost and 58 percent lower peak GPU memory usage.

Azercell Telecom LLC, Azerbaijan s leading telecom provider, teamed with AWS Generative AI Innovation Center to build an Azerbaijani large language model for telecom use cases and a customer facing chatbot. The challenge was real, morphologically rich Azerbaijani with limited training data and no blueprint for efficient LLM training. The teams set out to create a production ready framework on Amazon SageMaker AI that could scale in a real world setting and deliver measurable gains.

The effort yielded concrete engineering wins. By operating on an ml.p5.48xlarge instance and applying kernel level optimizations, the project achieved a 23 percent increase in training throughput and a 58 percent drop in peak GPU memory usage. In addition, a 2x improvement in tokens per word was attained through a custom tokenizer, effectively doubling the amount of Azerbaijani text that fits inside the model s context window. These improvements translate into faster experiments, tighter budgets, and the ability to train and test more iterations in the same time frame, a crucial edge when data is scarce or hard to collect.

The solution rests on a three stage framework that yields artifacts feeding the next stage. Stage 1 focuses on tokenizer development, evaluating three approaches: baseline English optimized tokenizers, vocabulary extension, and a custom monolingual tokenizer, and measuring encoding efficiency to select a practical path for a morphologically rich language. The pipeline emphasizes moving from tokenization choices to memory and throughput optimizations, then to model fine tuning and evaluation, all within a production ready workflow. The team notes that the tokenizer approach plays a pivotal role in how efficiently the model can learn from limited Azerbaijani data.

Open source tools underpinned the effort. The framework builds on PyTorch, Hugging Face Transformers, and Liger Kernels, with Azercell and AWS highlighting how established ML tooling can be adapted to language specific challenges. The collaboration also benefited from the AWS Generative AI Innovation Center, which helped shape a production ready pipeline on SageMaker AI designed for telecom workflows and a customer facing chatbot, indicating how cloud based AI platforms can shorten time to deployment for language models in real world products.

The project offers a number of practitioner takeaways for teams aiming to deploy language models for low resource or morphologically complex languages. First, tokenizer design matters. Testing diverse approaches, including monolingual tokenizers, should be part of any production plan because encoding efficiency directly affects memory footprint and the potential context window. Second, memory management at the kernel level can unlock sizable gains. The six week timeline and the 23 percent throughput jump show how targeted low level optimizations can dramatically shift the feasibility of experiments in a production setting. Third, a production ready pipeline on a managed platform like SageMaker AI can dramatically shorten the road from research to deployment, especially when paired with an open source stack that engineers already trust. Fourth, improving the tokens per word and the effective context window is not just a formatting touch; it changes how much text the model can reason over in a single pass, which matters for telecom chatbots and other customer interactions where coherence over longer prompts is valuable.

The Azercell AWS project demonstrates what a tightly scoped, production oriented collaboration can achieve for a language with limited data and complex morphology. It also signals a practical blueprint for other teams confronting similar constraints, showing that careful tokenizer design, kernel level optimizations, and a production ready SageMaker AI framework can yield meaningful gains in speed, memory usage, and text efficiency without waiting on theoretical breakthroughs.