This PR updates the plot_th_lat.py throughput-latency plot generation script to remove the concept of a backend (aml, fastgen, vllm) and generalize for any result output directory, irrespective of where it was run.
The PR also introduces the concept of an optionalplot_config.yaml that resides within each result directory and allows for overrides in the plot formatting. An example config file may look like this:
Each of the config parameters is optional, allowing for override of only the specific plot aspects required, however all parameters may also be provided.
A few nuances for the polyfit_degree and x/y_max parameters:
polyfit_degree: Specifies the polynomial degree for the 'best fit line'. Specifying 0 removes the best fit line and simply connects the scatter plot points.
x/y_max: Clips the x or y axis data using the specified value as the upper bound.
An example command executing the script may look something like this:
This PR updates the
plot_th_lat.py
throughput-latency plot generation script to remove the concept of abackend (aml, fastgen, vllm)
and generalize for any result output directory, irrespective of where it was run.The PR also introduces the concept of an optional
plot_config.yaml
that resides within each result directory and allows for overrides in the plot formatting. An example config file may look like this:Each of the config parameters is optional, allowing for override of only the specific plot aspects required, however all parameters may also be provided.
A few nuances for the
polyfit_degree
andx/y_max
parameters:polyfit_degree
: Specifies the polynomial degree for the 'best fit line'. Specifying0
removes the best fit line and simply connects the scatter plot points.x/y_max
: Clips the x or y axis data using the specified value as the upper bound.An example command executing the script may look something like this:
Or each result directory can be enumerated explicitly: