File: cache_debug_microbenchmarks.py

package info (click to toggle)
pytorch-cuda 2.6.0%2Bdfsg-7
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid, trixie
  • size: 161,620 kB
  • sloc: python: 1,278,832; cpp: 900,322; ansic: 82,710; asm: 7,754; java: 3,363; sh: 2,811; javascript: 2,443; makefile: 597; ruby: 195; xml: 84; objc: 68
file content (32 lines) | stat: -rw-r--r-- 558 bytes parent folder | download | duplicates (3)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
import timeit

import torch.fx
from torch._inductor.codecache import FxGraphHashDetails


N = 10000
K = 100


def huge_graph():
    def fn(x):
        for _ in range(N):
            x = x.sin()
        return x

    return torch.fx.symbolic_trace(fn)


def main():
    g = huge_graph()
    details = FxGraphHashDetails(g, [], {}, [])

    def fn():
        return details.debug_lines()

    t = min(timeit.repeat(fn, number=K, repeat=3))
    print(f"iterating over {N*K} FX nodes took {t:.1f}s ({N*K/t:.0f} nodes/s)")


if __name__ == "__main__":
    main()