Suggestions
Kevin Rohling
Kevin Rohling is a highly accomplished professional with a strong background in Artificial Intelligence and Computer Science.
He holds a Master of Science in Artificial Intelligence from The University of Texas at Austin and a Bachelor of Science in Computer Science from the University of Central Florida.
Kevin has held key positions at various organizations, including serving as the Head of Artificial Intelligence, Sr Engineering Lead, and VP of Engineering at Presence Product Group, LLC.
His expertise also extends to roles such as Founding Engineer/CTO at BookNook, AI Product Manager at Boomtrain, and Founding Engineer/VP Eng at emberlight inc.
Additionally, he has experience as a Technical Advisor, CEO, and CTO at different companies acquired by Electric Cloud.
Kevin Rohling's professional journey includes working at renowned companies like The Walt Disney Company in various roles such as Senior Software Engineer and Programmer Analyst.
His passion lies in building innovative solutions with AI that resonate with users.
Kevin's diverse experience and strong educational background make him a valuable asset in the fields of Artificial Intelligence and Engineering.
Highlights
Benchmarking Llama 2 70B on g5.12xlarge vs A100
We recently compiled inference benchmarks running upstage_Llama-2-70b-instruct-v2 on two different hardware setups. If you'd like to see the spreadsheet with the raw data you can check out this link: https://lnkd.in/gSSXY9Ra
Hardware Config #1: AWS g5.12xlarge - 4 x A10 w/ 96GB VRAM Hardware Config #2: Vultr - 1 x A100 w/ 80GB VRAM
A few questions we wanted to answer:
- How does the inference speed (tokens/s) between these two configurations compare?
- How does the number of input tokens impact inference speed?
- How many input tokens can these machines handle before they start to hit OOM?
- How does 4-bit vs 8-bit quantization affect all of the above?
Why this model? We chose upstage_Llama-2-70b-instruct-v2 because it's the current #1 performing OS model on HuggingFace's LLM Leaderboard. Also, according to the documentation the model is able to support 10K+ tokens using RoPE which is allowed us to push memory on the machines to the point of OOM.
Why this hardware? We have some projects that require high performance LLMs and these are the two most common hardware configurations that we've been considering. We do most of our cloud work on AWS so the g5.12xlarge is the "go to" option for inference with a model of this size. However, we've have been very interested in understanding if there are compelling reasons to go with a 1xA100 setup instead.
Text Generation Performance (t/s) vs Input Tokens (t) The chart shows how Text Generation Performance (t/s) responds to the number of input tokens (t) sent to the model. As expected, more input tokens results in slower generation speed.
GPU Load Performance (MM:SS) This is a measure of how long it took to load the model into memory. This was averaged this across 5 load attempts for each configuration.
Hardware: 8-Bit GPU Load Time, 4-Bit GPU Load Time g5.12xlarge (4xA10): 0:59, 1:00 1xA100: 2:47, 2:54
Average Text Generation Performance (tokens/second) Note that these numbers are an average across all text generation attempts for each configuration.
Hardware: 8-Bit (tokens/second), 4-Bit (tokens/second) g5.12xlarge (4xA10): 2.07 t/s, 4.08 t/s 1xA100: 2.28 t/s, 4.54 t/s
Maximum Context (tokens) This was a measure of how many input tokens we could pass into the model before getting an OOM exception for each configuration.
Hardware: 8-Bit Maximum Context (t), 4-Bit Maximum Context (t) g5.12xlarge (4xA10): 2500 tokens, 5500 tokens 1xA100: 3000 tokens, 8000 tokens
Summary On text generation performance the A100 config outperforms the A10 config by ~11%. The A100 config, which has less VRAM (80GB vs 96GB), was still able to handle a larger context size before hitting OOM errors. Unsurprisingly, 4-bit quantized models were much faster than 8-bit quantized models (almost 2x) and they were able to handle much larger context sizes before OOM.
