| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import math |
| 5 | import pickle |
| 6 | from collections import defaultdict |
| 7 | |
| 8 | import matplotlib.pyplot as plt |
| 9 | import pandas as pd |
| 10 | import regex as re |
| 11 | import seaborn as sns |
| 12 | from torch.utils.benchmark import Measurement as TMeasurement |
| 13 | |
| 14 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 15 | |
| 16 | if __name__ == "__main__": |
| 17 | parser = FlexibleArgumentParser( |
| 18 | description="Benchmark the latency of processing a single batch of " |
| 19 | "requests till completion." |
| 20 | ) |
| 21 | parser.add_argument("filename", type=str) |
| 22 | |
| 23 | args = parser.parse_args() |
| 24 | |
| 25 | with open(args.filename, "rb") as f: |
| 26 | data = pickle.load(f) |
| 27 | raw_results: list[TMeasurement] = data["results"] |
| 28 | |
| 29 | results = defaultdict(lambda: list()) |
| 30 | for v in raw_results: |
| 31 | result = re.search(r"MKN=\(\d+x(\d+x\d+)\)", v.task_spec.sub_label) |
| 32 | if result is not None: |
| 33 | KN = result.group(1) |
| 34 | else: |
| 35 | raise Exception("MKN not found") |
| 36 | result = re.search(r"MKN=\((\d+)x\d+x\d+\)", v.task_spec.sub_label) |
| 37 | if result is not None: |
| 38 | M = result.group(1) |
| 39 | else: |
| 40 | raise Exception("MKN not found") |
| 41 | |
| 42 | kernel = v.task_spec.description |
| 43 | results[KN].append({"kernel": kernel, "batch_size": M, "median": v.median}) |
| 44 | |
| 45 | rows = int(math.ceil(len(results) / 2)) |
| 46 | fig, axs = plt.subplots(rows, 2, figsize=(12, 5 * rows)) |
| 47 | axs = axs.flatten() |
| 48 | for axs_idx, (shape, data) in enumerate(results.items()): |
| 49 | plt.sca(axs[axs_idx]) |
| 50 | df = pd.DataFrame(data) |
| 51 | sns.lineplot( |
| 52 | data=df, |
| 53 | x="batch_size", |
| 54 | y="median", |
| 55 | hue="kernel", |
| 56 | style="kernel", |
| 57 | markers=True, |
| 58 | dashes=False, |
| 59 | palette="Dark2", |
| 60 | ) |
| 61 | plt.title(f"Shape: {shape}") |
| 62 | plt.ylabel("time (median, s)") |
| 63 | plt.tight_layout() |
| 64 | plt.savefig("graph_machete_bench.pdf") |