zenowzh/llama2_run

llama2_run's Issues

Assumptions to explore

• training time linearity and the relationship between data set sizes
• training time linearity and the relationship between model parameter sizes
• training time linearity and the relationship between batch sizes (?)

or other parameters we can use to potentially extrapolate single GPU benchmark results to models of the same class, and eventually make useful assumptions for multi GPU setup.

e.g. is it possible to make the following conclusion with certain setups?
a 10x in training data set size results in about X times increase in training time for transformer based models under this circumstance

we can start by graphing our existing benchmark results to compare the data points.

reference GPUs to run tests on

• H100 80GB
• A100 80GB
• L40 48GB
• A6000 48GB
• A5000 24GB
• A4000 16GB
• A40 48GB
• A30 24GB

Model to test

Here are the models for benchmark testing, better check the license :

• LLMs

• LLama2_7b
• LLama2_13b
• Falcon_40b
• LLama2_70b

• Text_to_images