scripts/plot_shape_distributions.py

"""Render one exact-shape histogram per definition from workload metadata only."""

import argparse
from collections import Counter
import json
import os
from pathlib import Path
import textwrap


ROOT = Path(__file__).resolve().parents[1]
os.environ.setdefault("MPLCONFIGDIR", str(ROOT / "results/shape-distributions/matplotlib"))

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.ticker import MaxNLocator


def load_distributions():
    distributions = []
    seen_uuids = set()
    for path in sorted((ROOT / "definitions").rglob("*.json")):
        definition = json.loads(path.read_text())
        workload_path = (ROOT / "workloads" / path.relative_to(ROOT / "definitions")).with_suffix(".jsonl")
        axes = [name for name, spec in definition["axes"].items() if spec["type"] == "var"]
        counts = Counter()
        for line in workload_path.read_text().splitlines():
            if not line.strip():
                continue
            row = json.loads(line)
            workload = row["workload"]
            if row["definition"] != definition["name"] or set(workload["axes"]) != set(axes):
                raise ValueError(f"Definition or runtime dimensions mismatch: {workload_path}")
            if workload["uuid"] in seen_uuids:
                raise ValueError(f"Duplicate workload UUID: {workload['uuid']}")
            seen_uuids.add(workload["uuid"])
            counts[tuple(workload["axes"][axis] for axis in axes)] += 1
        if not counts:
            raise ValueError(f"No workloads: {workload_path}")
        distributions.append((path, workload_path, definition, axes, counts))
    return distributions


def plot_distribution(definition, axes, counts):
    shapes = sorted(counts)
    frequencies = [counts[shape] for shape in shapes]
    horizontal = len(axes) != 1
    labels = [str(shape[0]) if len(axes) == 1 else str(shape) for shape in shapes]
    height = max(4.8, 2.7 + len(shapes) * 0.28) if horizontal else 5.4
    fig = plt.figure(figsize=(11.8, height), facecolor="white")
    left = max(2.0, max(map(len, labels)) * 0.077 + 0.25) if horizontal else 0.8
    bottom = 0.9 if horizontal else 1.12
    top = 1.85 if horizontal else 1.55
    ax = fig.add_axes((left / 11.8, bottom / height, (11.35 - left) / 11.8,
                       (height - bottom - top) / height))
    color = "#2563a6"
    fig.text(0.035, 1 - 0.27 / height, "Shape distribution", fontsize=16, weight="bold", color="#142c44")
    fig.text(0.035, 1 - 0.52 / height, definition["name"], fontsize=10, family="monospace", color="#334155")
    fig.text(0.035, 1 - 0.78 / height,
             f"{sum(frequencies)} captured workloads  |  {len(shapes)} distinct shapes",
             fontsize=11, color=color, weight="bold")
    fixed = ", ".join(f"{name}={spec['value']}" for name, spec in definition["axes"].items()
                      if spec["type"] == "const")
    fig.text(0.035, 1 - 1.02 / height, textwrap.fill("Fixed dimensions: " + fixed, width=125),
             fontsize=9, va="top", color="#526171")
    if horizontal:
        bars = ax.barh(range(len(shapes)), frequencies, height=0.68, color=color)
        ax.set_yticks(range(len(shapes)), labels, fontfamily="monospace", fontsize=9)
        ax.invert_yaxis()
        ax.set_xlabel("Captured workloads (count)", labelpad=10)
        fig.text(left / 11.8, 1 - 1.55 / height, "Runtime shape: (" + ", ".join(axes) + ")",
                 fontsize=10, weight="bold", color="#334155")
        ax.set_xlim(0, max(frequencies) * 1.13)
        ax.xaxis.set_major_locator(MaxNLocator(integer=True, nbins=8))
        ax.grid(axis="x", color="#e3e9ef", linewidth=0.8)
    else:
        bars = ax.bar(range(len(shapes)), frequencies, width=0.72, color=color)
        ax.set_xticks(range(len(shapes)), labels, rotation=35, ha="right", fontsize=10)
        description = definition["axes"][axes[0]].get("description", axes[0]).rstrip(".")
        ax.set_xlabel(f"{axes[0]} — {description}", labelpad=12)
        ax.set_ylabel("Captured workloads (count)", labelpad=10)
        ax.set_ylim(0, max(frequencies) * 1.17)
        ax.yaxis.set_major_locator(MaxNLocator(integer=True, nbins=7))
        ax.grid(axis="y", color="#e3e9ef", linewidth=0.8)
    ax.bar_label(bars, padding=4, fontsize=9, color="#243c52")
    ax.set_axisbelow(True)
    ax.spines[["top", "right"]].set_visible(False)
    ax.spines[["left", "bottom"]].set_color("#b8c5d1")
    ax.tick_params(axis="both", length=0, pad=7, colors="#334155")
    fig.text(0.035, 0.10 / height,
             "One bin per exact runtime shape; equal spacing. Counts describe this dataset, not production traffic.",
             fontsize=8.5, color="#526171")
    return fig


def write_gallery(distributions, output):
    total = sum(sum(counts.values()) for _, _, _, _, counts in distributions)
    lines = [
        "# Kernel shape distributions", "",
        f"**{len(distributions)} kernels, {total} captured workloads.** Each histogram counts the workload records "
        "in one kernel definition's JSONL file. Every bar is one exact combination of its runtime dimensions; "
        "identical shapes are combined. The labels retain all runtime dimensions, including dimensions that "
        "happen to have one observed value. Model constants appear above each plot.", "",
        "Read bar height (or length in horizontal plots) as the number of captured workloads with that shape. "
        "Bins are equally spaced categories, sorted by the dimension values, not continuous numeric intervals. "
        "A bar of 1 means one published example; a bar of 6 means six. These are capture counts, "
        "not request probabilities, runtime invocation frequencies, or AgentX traffic weights.", "",
        "Multi-dimensional shapes use horizontal bars to keep every dimension readable. "
        "Dimension names are explained below each plot; the linked definition specifies complete tensor layouts.", "",
        f"Regenerate from the repository root with Python and Matplotlib (rendered with {matplotlib.__version__}):", "",
        "```bash", "python scripts/plot_shape_distributions.py", "```", "",
        "Only definition and workload metadata are needed; tensor snapshots and a GPU are unnecessary.", "",
        "| Kernel definition | Workloads | Distinct shapes |", "|---|---:|---:|",
    ]
    for _, _, definition, _, counts in distributions:
        name = definition["name"]
        lines.append(f"| [{name}](#{name}) | {sum(counts.values())} | {len(counts)} |")
    for path, workload_path, definition, axes, counts in distributions:
        name = definition["name"]
        lines.extend([
            "", f"## {name}", "",
            f"[Definition](../../{path.relative_to(ROOT)}) · "
            f"[Workloads](../../{workload_path.relative_to(ROOT)}) · "
            f"[SVG]({name}.svg)", "",
            f"![Shape distribution for {name}: {sum(counts.values())} workloads across {len(counts)} shapes]({name}.svg)", "",
            "**Runtime dimensions:** " + " ".join(
                f"`{axis}`: {definition['axes'][axis].get('description', axis)}" for axis in axes),
        ])
    (output / "README.md").write_text("\n".join(lines) + "\n")


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--preview-dir", type=Path, help="Optionally save PNG previews outside the published gallery.")
    args = parser.parse_args()
    distributions = load_distributions()
    output = ROOT / "docs/shape-distributions"
    output.mkdir(parents=True, exist_ok=True)
    if args.preview_dir:
        args.preview_dir.mkdir(parents=True, exist_ok=True)
    plt.rcParams.update({"font.family": "DejaVu Sans", "svg.fonttype": "none", "svg.hashsalt": "kernel-shapes"})
    for _, _, definition, axes, counts in distributions:
        fig = plot_distribution(definition, axes, counts)
        name = definition["name"]
        svg_path = output / f"{name}.svg"
        fig.savefig(svg_path, metadata={"Date": None, "Title": f"Shape distribution: {name}"})
        # Matplotlib adds trailing spaces to path coordinates; keep generated diffs clean.
        svg_path.write_text("\n".join(line.rstrip() for line in svg_path.read_text().splitlines()) + "\n")
        if args.preview_dir:
            fig.savefig(args.preview_dir / f"{name}.png", dpi=120)
        plt.close(fig)
    write_gallery(distributions, output)
    total = sum(sum(counts.values()) for _, _, _, _, counts in distributions)
    print(f"Wrote {len(distributions)} histograms covering {total} workloads to {output}")


if __name__ == "__main__":
    main()