Visualization¶

The ipw plot command generates visualizations from profiling results. IPW ships with two visualization providers: regression (scatter plots with regression lines) and output-kde (output length distributions).

Command Reference¶

ipw plot <results_dir> [--output <dir>]

Option	Default	Description
`<results_dir>`	required	Path to the profiling results directory
`--output`	`<results_dir>/plots/`	Output directory for plots

Built-in Visualizations¶

Regression Scatter Plots¶

Generates scatter plots of energy, power, and latency against input and output token counts, with fitted regression lines:

Energy (J) vs. Input Tokens
Energy (J) vs. Output Tokens
Power (W) vs. Input Tokens
Latency (s) vs. Output Tokens

Each plot shows individual data points and the best-fit line, making it easy to see how resource consumption scales with workload.

Output KDE¶

Generates a kernel density estimate (KDE) plot showing the distribution of output token lengths across all queries. This helps identify whether the model tends to produce short or long responses for the given dataset.

Output¶

Plots are saved as PNG files in the output directory:

runs/profile_<hardware>_<model>/
    plots/
        regression.png
        output_kde.png

Custom Visualization Providers¶

Create custom visualizations by subclassing VisualizationProvider and registering:

from ipw.visualization.base import VisualizationProvider
from ipw.core.registry import VisualizationRegistry


@VisualizationRegistry.register("my-viz")
class MyVisualization(VisualizationProvider):
    visualization_id = "my-viz"

    def run(self, results_dir, output_dir=None):
        # Load data from results_dir
        # Generate plots with matplotlib
        # Save to output_dir
        ...

Listing Available Visualizations¶

ipw list visualizations