Caroline Bishop
November 22, 2024 01:19
NVIDIA’s TensorRT-LLM solves the problem of long sequence lengths by introducing multi-block attention to dramatically improve AI inference throughput by up to 3.5x on HGX H200.
In a significant development for AI inference, NVIDIA has unveiled the TensorRT-LLM multi-block attention feature that significantly improves throughput on the NVIDIA HGX H200 platform. According to NVIDIA, this innovation addresses the growing needs of modern generative AI models by improving throughput by more than 3x for long sequence lengths.
Advances in Generative AI
The rapid advancement of generative AI models, exemplified by the Llama 2 and Llama 3.1 series, has introduced models with much larger context windows. For example, the Llama 3.1 model supports context lengths of up to 128,000 tokens. While this expansion allows AI models to perform complex cognitive tasks on a wide range of data sets, it also presents unique challenges in the AI inference environment.
Challenges of AI inference
AI inference, especially with long sequence lengths, faces obstacles such as low latency requirements and small batch size requirements. Existing GPU deployment methods often do not properly utilize the streaming multiprocessor (SM) of NVIDIA GPUs, especially during the decoding phase of inference. This lack of utilization impacts overall system throughput. This is because only a small portion of the GPU SM is used, leaving many resources idle.
Multi-block attention solution
NVIDIA’s TensorRT-LLM multiblock attention solves this challenge by maximizing GPU resource usage. Divide the computation task into smaller blocks and distribute them to all available SMs. This not only alleviates memory bandwidth limitations, but also improves throughput by efficiently utilizing GPU resources during the decoding phase.
Performance of NVIDIA HGX H200
NVIDIA HGX H200’s multi-block attention implementation showed surprising results. This allows the system to generate up to 3.5x more tokens per second for long sequence queries in low-latency scenarios. Using model parallelism, a 3x performance improvement is observed without affecting the time to first token, even when half the GPU resources are used.
Implications and future prospects
These advances in AI inference technology allow existing systems to support longer context lengths without additional hardware investments. TensorRT-LLM multi-block attention is enabled by default, significantly improving the performance of AI models with extensive context requirements. This development highlights NVIDIA’s commitment to advancing AI inference capabilities to more efficiently process complex AI models.
Image source: Shutterstock