Box 10.1.0.201, GPU1, registry.arbi.work/arbi-serve:test-e54f6a4 image
(same torch/exllamav3 pins as the production boot), no model loaded — pure
torch.library.custom_op mechanism test. Script:
graph_capture_dispatch_microbench.py (this directory).

    EAGER: 200 iters x 184 calls = 36800 total op calls
    EAGER: backend_impl tottime = 0.6071s of 1.3944s profiled (43.5%)
    REPLAY: 200 graph.replay() calls (each replaying 184 baked op calls)
    REPLAY: backend_impl tottime = 0.0000s of 0.0940s profiled (0.0%)

    === SUMMARY ===
    eager backend_impl fraction:  43.5%
    replay backend_impl fraction: 0.0%
    replay total profiled time for 200 chunk-equivalent replays: 0.0940s (0.470 ms/replay)
    eager total profiled time for 200 chunk-equivalent iters: 1.3944s (6.972 ms/iter)
