| |
| """Render the three recorded V4.1 sweeps in the Qwen model-card style.""" |
| import csv |
| import json |
| import math |
| from pathlib import Path |
| from statistics import mean, median |
|
|
| import matplotlib |
|
|
| matplotlib.use("Agg") |
| import matplotlib.pyplot as plt |
|
|
|
|
| DATA = Path(__file__).with_name("ds41-spark-64k-2026-10-13") |
| FRONTIERS = list(range(2048, 65537, 2048)) |
| REPEATS = 3 |
| GENERATED = 128 |
| COLOR = "#087f8c" |
| METRICS = ( |
| ("prefill_tps", "Full prompt prefill"), |
| ("gen_tps", "Decode (automatic MTP policy)"), |
| ) |
|
|
|
|
| def load_runs(): |
| runs = [] |
| for repeat in range(1, REPEATS + 1): |
| with (DATA / f"run-{repeat}.csv").open() as source: |
| rows = list(csv.DictReader(source)) |
| assert [int(row["ctx_tokens"]) for row in rows] == FRONTIERS |
| assert all(int(row["gen_tokens"]) == GENERATED for row in rows) |
| assert all(int(row["prefill_cached_tokens"]) == 0 for row in rows) |
| for key, _ in METRICS: |
| assert all(math.isfinite(float(row[key])) and float(row[key]) > 0 for row in rows) |
| runs.append(rows) |
| inputs = [[int(row["prefill_tokens"]) for row in run] for run in runs] |
| assert all(row == inputs[0] for row in inputs) |
| return runs |
|
|
|
|
| def summarize(runs): |
| report = {"repeats": REPEATS, "frontiers": [], "metrics": {}} |
| for key, _ in METRICS: |
| run_means = [mean(float(row[key]) for row in run) for run in runs] |
| report["metrics"][key] = { |
| "run_means": run_means, |
| "median_run_mean": median(run_means), |
| } |
| for index, frontier in enumerate(FRONTIERS): |
| row = {"frontier_tokens": frontier, |
| "actual_prompt_tokens": int(runs[0][index]["prefill_tokens"]), |
| "generated_tokens": GENERATED} |
| for key, _ in METRICS: |
| values = [float(run[index][key]) for run in runs] |
| row[key] = {"runs": values, "median": median(values), |
| "min": min(values), "max": max(values)} |
| report["frontiers"].append(row) |
| return report |
|
|
|
|
| def write_tables(report): |
| lines = ["# DeepSeek V4.1 Flash: 2K–64K throughput", "", |
| "Tokens/second on one DGX Spark / GB10 at 300–2200 MHz.", |
| "Three fresh server processes; all samples retained. Each request has", |
| "zero reused tokens and 128 generated tokens. Frontiers come from the", |
| "corpus byte calibration; the input column is the actual token count.", "", |
| "[Conditions, graph and evidence](README.md).", ""] |
| for key, heading in METRICS: |
| lines.extend([f"## {heading}", "", |
| "| Frontier | Input tokens | Run 1 | Run 2 | Run 3 | Median | Min–max |", |
| "|---:|---:|---:|---:|---:|---:|---:|"]) |
| for row in report["frontiers"]: |
| value = row[key] |
| samples = " | ".join(f"{sample:,.1f}" for sample in value["runs"]) |
| lines.append(f"| {row['frontier_tokens'] // 1024}K | {row['actual_prompt_tokens']:,} | " |
| f"{samples} | **{value['median']:,.1f}** | " |
| f"{value['min']:,.1f}–{value['max']:,.1f} |") |
| summary = report["metrics"][key] |
| lines.extend(["", f"Median of the three run means: **{summary['median_run_mean']:,.2f} tok/s**.", ""]) |
| (DATA / "tables.md").write_text("\n".join(lines)) |
|
|
|
|
| def plot(report): |
| plt.rcParams.update({"font.size": 11, "axes.spines.top": False, |
| "axes.spines.right": False, "savefig.facecolor": "white"}) |
| fig, axes = plt.subplots(2, 1, figsize=(11, 8), sharex=True) |
| fig.suptitle("DeepSeek V4.1 Flash · VQ + code GR", fontsize=18, |
| fontweight="bold", y=0.98) |
| fig.text(0.5, 0.93, "One DGX Spark / GB10 · 2K–64K full-prefill sweep · 2026-10-13", |
| ha="center", color="#4b5563") |
| x = [row["frontier_tokens"] / 1024 for row in report["frontiers"]] |
| for ax, (key, heading) in zip(axes, METRICS): |
| values = [row[key] for row in report["frontiers"]] |
| center = report["metrics"][key]["median_run_mean"] |
| ax.fill_between(x, [row["min"] for row in values], [row["max"] for row in values], |
| color=COLOR, alpha=0.17, linewidth=0) |
| ax.plot(x, [row["median"] for row in values], color=COLOR, linewidth=2, |
| label=f"VQ + code GR · {center:,.1f} tok/s") |
| ax.set_title(heading, loc="left", fontsize=13, pad=10) |
| ax.set_ylabel("Tokens / second") |
| ax.grid(alpha=0.2) |
| ax.legend(loc="lower right", framealpha=0.92) |
| ax.margins(y=0.2) |
| axes[-1].set_xlabel("Context frontier (K = 1,024)") |
| axes[-1].set_xticks([2, 8, 16, 24, 32, 40, 48, 56, 64]) |
| axes[-1].set_xlim(2, 64) |
| fig.text(0.5, 0.06, |
| "Curves: per-frontier medians; bands: min–max of 3 fresh runs; legend: median of run means.\n" |
| "Full prompt prefill + 128 greedy tokens · zero KV reuse · actual input counts in the report\n" |
| "Original Engram + code GR · automatic MTP · CUDA graphs · 300–2200 MHz\n" |
| "DeepSeek V4.1 Flash ported to ds4-dfm-rs by JLFN (PR #80) · YoungAi VQ artifacts", |
| ha="center", va="center", fontsize=9, color="#4b5563", linespacing=1.6) |
| fig.subplots_adjust(top=0.87, bottom=0.17, left=0.1, right=0.97, hspace=0.3) |
| fig.savefig(DATA / "throughput.png", dpi=180) |
| plt.close(fig) |
|
|
|
|
| if __name__ == "__main__": |
| report = summarize(load_runs()) |
| (DATA / "summary.json").write_text(json.dumps(report, indent=2) + "\n") |
| write_tables(report) |
| plot(report) |
| print(json.dumps(report["metrics"], indent=2)) |
|
|
