#!/usr/bin/env python3 """Write the DeepSpeed/analytic estimator supplement section from results/.""" from __future__ import annotations import argparse import csv import json from pathlib import Path HERE = Path(__file__).resolve().parent METHODS = [ ("ds3_full", "ZeRO-3, all parameters trainable"), ("ds3_lora+frozen_base", "ZeRO-3, LoRA states + frozen BF16 base"), ("ds3_lora+frozen_base+reservation", "\\quad + vLLM reservation"), ("ds3_lora+frozen_base+reservation+activations", "\\quad + reservation + activations"), ("ds2_full", "ZeRO-2, all parameters trainable"), ("analytic_bf16", "Analytic component budget (BF16)"), ("source_label", "Source-label copy (reference)"), ("always_approve", "Always approve (reference)"), ] def main() -> None: parser = argparse.ArgumentParser() parser.add_argument("--results", type=Path, default=HERE / "results") parser.add_argument("--output", type=Path, required=True) args = parser.parse_args() with (args.results / "score_comparison_all_methods.csv").open() as handle: scores = {row["method"]: row for row in csv.DictReader(handle) if row["track"] == "all"} summary = json.loads((args.results / "deepspeed_summary.json").read_text()) rows = [] for key, name in METHODS: row = scores[key] rows.append(f"{name} & {float(row['three_state_accuracy']):.3f} & {float(row['within_margin_recall']):.3f} & " f"{row['approved_memory_failure']}/{row['target_memory_failure']} & " f"{row['approved_above_margin']}/{row['target_above_margin']} \\\\") any_row = next(iter(scores.values())) text = [ r"\section{Query-level comparison with DeepSpeed memory estimators}\label{ssec:deepspeed}", "", "This comparison complements the native LLMem validation. It applies DeepSpeed's published " f"model-state estimators (DeepSpeed {summary['deepspeed_version']}, called directly; " r"\nolinkurl{estimate_zero3_model_states_mem_needs} and its ZeRO-2 counterpart) to the " f"{any_row['queries']} public benchmark queries, using only target settings and model architecture. " "These queries reuse 94 distinct target configurations, so they are not independent experiments. " "Predicted per-device memory is compared with the operating limit and capacity to give a three-state " "label, scored with the released evaluator. Composites add the requested vLLM reservation and an " "activation term. A transparent analytic component budget is shown for reference. These estimators " "target training state; applying them to colocated RL is outside their intended scope, so the " "comparison shows which terms they omit, not their accuracy for training.", "", r"\begin{table}[!ht]\centering\small", r"\caption{Estimator decisions on all benchmark targets. Recall is within-margin recall; approval means a within-margin prediction; failures and above-margin approvals are counted against all such targets.}", r"\label{stab:deepspeed}", r"\begin{tabular}{@{}lrrrr@{}}\toprule", r"Predictor & Acc. & Recall & Appr.\ failures & Appr.\ above \\\midrule", *rows, r"\bottomrule\end{tabular}\end{table}", "", "Model states alone approve every memory-failing target. Adding the reservation removes few of these " "approvals, and adding activations rejects many usable targets because stages that never coexist " "under state release are summed. The released artifact (\\texttt{estimator\\_comparison/}) records " "every estimator call with its arguments, returned bytes, and source hashes.", "", ] args.output.write_text("\n".join(text) + "\n") if __name__ == "__main__": main()