Lao Huang gives H100 a boost: Nvidia launches large model acceleration package, doubling Llama2 inference speed

The inference speed of large models has doubled in just one month!

Recently, Nvidia announced the launch of a "chicken blood package" specially designed for H100, aiming to speed up the LLM inference process

Maybe now you don't have to wait for the GH200 to be delivered next year. .

The computing power of GPU has been affecting the performance of large models. Both hardware suppliers and users hope to obtain faster computing speed

As the largest supplier of hardware behind large models, NVIDIA has been studying how to accelerate large model hardware.

Through cooperation with a number of AI companies, NVIDIA finally launched the large model inference optimization program TensorRT-LLM (tentatively referred to as TensorRT).

TensorRT can not only double the inference speed of large models, but is also very convenient to use.

Without in-depth knowledge of C and CUDA, you can quickly customize optimization strategies and run large models faster on H100.

NVIDIA scientist Jim Fan forwarded and commented that NVIDIA’s “another advantage” is the supporting software that can maximize the use of GPU performance.

NVIDIA injects new vitality into its products through software, just like it implements Lao Huang's saying "the more you buy, the more you save." However, this does not prevent some people from thinking that the price of the product is too high

In addition to the price, some netizens also questioned its operating effect:

We always I have seen how many times the performance has improved (as advertised), but when I run Llama 2 myself, I can still only process dozens of tokens per second.

For TensorRT, we need further testing to determine whether it is really effective. Let us first take a closer look at TensorRT

Double the inference speed of large models

TensorRT-LLM optimized H100, how fast is it for running large models?

Nvidia’s announcement provides data for two models, Llama 2 and GPT-J-6B.

On the optimized H100, the inference speed of running Llama 2 is 4.6 times that of the A100 and 1.77 times that of the unoptimized H100 in August

The inference speed of GPT-J-6B is 8 times that of A100 and 2 times that of the August unoptimized version.

TensorRT also provides an open source modular Python API that can quickly customize optimization solutions according to different LLM requirements

This API will combine the deep learning compiler with , kernel optimization, pre/post-processing and multi-node communication functions are integrated together.

There are also customized versions for common models such as GPT(2/3) and Llama, which can be "used out of the box".

Through the latest open source AI kernel in TensorRT, developers can also optimize the model itself, including the attention algorithm FlashAttention that greatly speeds up Transformer.

TensorRT is a high-performance inference engine for optimizing deep learning inference. It optimizes LLM inference speed by using technologies such as mixed-precision computing, dynamic graph optimization, and layer fusion. Specifically, TensorRT improves inference speed by reducing the amount of computation and memory bandwidth requirements by converting floating-point calculations into half-precision floating-point calculations. In addition, TensorRT also uses dynamic graph optimization technology to dynamically select the optimal network structure based on the characteristics of the input data, further improving the inference speed. In addition, TensorRT also uses layer fusion technology to merge multiple computing layers into a more efficient computing layer, reducing computing and memory access overhead and further improving inference speed. In short, TensorRT has significantly improved the speed and efficiency of LLM inference through a variety of optimization technologies

First of all, it is due to TensorRT's optimization of multi-node collaborative working.

A huge model like Llama cannot be run on a single card. It requires multiple GPUs to run together.

In the past, this work required people to manually disassemble the model to achieve it.

With TensorRT, the system can automatically split the model and run it efficiently between multiple GPUs through NVLink

Secondly, TensorRT also An optimized scheduling technology called Dynamic Batch Processing is used.

During the inference process, LLM is actually performed by executing model iterations multiple times

Dynamic batch processing technology will kick out the completed sequence immediately instead of waiting for the entire batch of tasks Once complete, process the next set of requests.

In actual tests, dynamic batch processing technology successfully reduced LLM's GPU request throughput by half, thereby significantly reducing operating costs

Another key point is Convert 16-bit precision floating point numbers to 8-bit precision , thereby reducing memory consumption.

Compared with FP16 in the training phase, FP8 has lower resource consumption and is more accurate than INT-8. It can improve performance without affecting the accuracy of the model

Usage Hopper Transformer engine, the system will automatically complete the conversion and compilation of FP16 to FP8, without the need to manually modify any code in the model

Currently, the early bird version of TensorRT-LLM is available for download, and the official version will be launched in a few weeks And integrated into the NeMo framework

One More Thing

Whenever a big event occurs, the figure of "Leewenhoek" is indispensable.

In Nvidia’s announcement, it mentioned cooperation with leading artificial intelligence companies such as Meta, but did not mention OpenAI

From this announcement, some netizens discovered this point and sent it to On the OpenAI forum:

Please let me see who has not been cueed by Lao Huang (manual dog head)

Are you still What kind of "surprises" do we expect Lao Huang to bring us?

The above is the detailed content of Lao Huang gives H100 a boost: Nvidia launches large model acceleration package, doubling Llama2 inference speed. For more information, please follow other related articles on the PHP Chinese website!