| """Static figure from observed training/quantization measurements.""" |
| from pathlib import Path |
| import json |
| import matplotlib |
| matplotlib.use("Agg") |
| import matplotlib.pyplot as plt |
|
|
|
|
| def main(): |
| r = json.loads(Path("artifacts/tiny/training-report.json").read_text()) |
| q = json.loads(Path("reports/quantization.json").read_text()) |
| plt.rcParams.update({"font.size": 10, "axes.spines.top": False, "axes.spines.right": False}) |
| fig, axes = plt.subplots(1, 2, figsize=(11, 4.3), layout="constrained") |
| rows = r["metrics"] |
| axes[0].plot([x["step"] for x in rows], [x["train_loss"] for x in rows], "o-", label="Training batch", color="#2456a6") |
| axes[0].plot([x["step"] for x in rows], [x["validation_loss"] for x in rows], "s-", label="Fixed validation batch", color="#c25823") |
| axes[0].set(xlabel="Optimization step", ylabel="Cross-entropy (nats / byte token)", title="820,736-parameter training experiment") |
| axes[0].legend(frameon=False) |
| names = ["fp32", "dynamic_int8_linear_only"] |
| p50 = [q[n]["batch_latency_seconds"]["p50"]*1000 for n in names] |
| p95 = [q[n]["batch_latency_seconds"]["p95"]*1000 for n in names] |
| x = [0, 1] |
| axes[1].bar([i-.17 for i in x], p50, width=.34, label="P50", color="#2456a6") |
| axes[1].bar([i+.17 for i in x], p95, width=.34, label="P95", color="#87a5d4") |
| axes[1].set_xticks(x, ["FP32", "INT8 linear layers"]) |
| axes[1].set(ylabel="CPU batch latency (ms)", title="20 samples; no clear tail-latency win") |
| axes[1].legend(frameon=False) |
| fig.suptitle("NEXORA local prototype: measured mechanics, not capability benchmarks", fontsize=12, fontweight="bold") |
| fig.savefig("reports/measurements.png", dpi=160) |
| plt.close(fig) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
