NVIDIA improves Llama 3.3 70B model performance with TensorRT-LLM

Meta’s latest addition to the Llama collection, the Llama 3.3 70B model, features significant performance improvements thanks to NVIDIA’s TensorRT-LLM. According to NVIDIA, the goal of this collaboration is to optimize the inference throughput of large language models (LLMs), increasing it by up to three times.

Advanced optimization with TensorRT-LLM

NVIDIA TensorRT-LLM uses several innovative technologies to maximize the performance of Llama 3.3 70B. Key optimizations include in-flight batching, KV caching, and custom FP8 quantization. These technologies are designed to improve LLM service efficiency, reduce latency, and improve GPU utilization.

Ongoing batch processing allows you to optimize throughput by processing multiple requests simultaneously. By interleaving requests across context and creation phases, we minimize latency and improve GPU utilization. Additionally, the KV cache mechanism saves computational resources by storing key-value elements of previous tokens, although it requires careful management of memory resources.

Speculative decoding technology

Speculative decoding is a powerful way to accelerate LLM inference. This allows us to generate multiple sequences of future tokens, which are processed more efficiently than a single token in autoregressive decoding. TensorRT-LLM supports a variety of speculative decoding techniques, including draft target, Medusa, Eagle, and predictive decoding.

These techniques significantly improve throughput, as evidenced by internal measurements using NVIDIA’s H200 Tensor Core GPUs. For example, using the draft model, throughput increases from 51.14 tokens per second to 181.74 tokens per second, achieving a 3.55x speedup.

Implementation and Deployment

To achieve these performance gains, NVIDIA provides a comprehensive setup to integrate the Llama 3.3 70B model with draft target speculative decoding. This includes downloading model checkpoints, installing TensorRT-LLM, and compiling model checkpoints with the optimized TensorRT engine.

NVIDIA’s commitment to advancing AI technology extends to collaborations with Meta and other partners aimed at advancing open community AI models. TensorRT-LLM optimizations not only improve throughput, but also reduce energy costs and improve total cost of ownership, making AI deployments more efficient across diverse infrastructures.

For more information about the setup process and further optimizations, visit the official NVIDIA blog.

Image source: Shutterstock