Who are you?

You are a passionate and driven individual who strives to be their best everyday. You have extensive experience building infrastructures that enable training, fine-tuning, and serving billion-parameter scale deep learning models, especially in the NLP domain using Pytorch and the huggingface ecosystem.

What you’ll be doing

• Build, train, and fine-tune large language models using Pytorch on advanced hardware setups like GPUs and TPUs, employing CUDA in multi-node, multi-GPU environments.
• Develop robust serving APIs to deliver sub-second latency inference for large language models, utilizing various optimization techniques.
• Continuously improve model performance by fine-tuning LLMs, embedding models, and rankers to meet specific application needs.

What should you have?

• 2+ years of experience working with large-scale Pytorch-based deep learning applications on GPUs and TPUs using CUDA in multi-node multi-GPU scenarios
• 2+ years of experience building, training and fine-tuning pipelines for large language models (LLMs) using distributed training approaches for both model and data
• 2+ years of experience building serving APIs for sub-second latency inference of large language models using various optimization techniques
• Extensive experience with Pytroch, DeepSpeed, Megatron-LM, and the Huggingface ecosystem
• Proven track record in fine-tuning embedding models, and re-rankers.
• 1+ years of experience with model training optimization and distributed training, using libraries such as HF Accelerate, BitsandBytes, and Flash-Attention to enhance training efficiency and scalability.
• Experience in managing machine learning experiments and monitoring model performance, using tools like Weights and Biases and MLFlow.
• Familiarity with embedding models inference libraries (e.g.,Infinity, Text Embedding Interface) and LLM inference libraries (e.g., VLLM, Text Generation Interface).

• Experience with Keras, JAX/FLAX – an advantage
• Experience with advanced finetuning methods such as RLHF, DPO, KTO, ORPO, etc.. – an advantage
• Experience with parameter efficient finetuning methods such as LoRA, DoRA, ReFT, IA3, etc.. – an advantage