# perfil-glm53-20260828

Public raw data package for this measurement. Every number on the measurements page citing this window comes from these files.

Summary derived from torch.profiler trace of rank 7 of 8 (tensor parallel 8), collected on 2026-08-28. The full trace is not published: 347,011,682 bytes, sha256 in field `trace_sha256`. Method is in field `metodo`.

Regime: single request, unspeculated decode, a 524-token prompt (the word 'alpha' repeated 512 times, chat template), greedy decoding (temperature 0), 64 output tokens, thinking disabled. The profiler elongates the step; timings serve to break down the step, not as throughput.

All-reduce message size (block `mensagem_do_all_reduce`) is derived from the model's public config, not measured in trace.

| file | what it is | sha256 (first 16) |
|---|---|---|
| `allreduce_posicao.json` | distribution of the 91 all-reduce calls by step position (measured under profiler) | `250e95e8d90f01b8` |
| `resumo.json` | means, standard deviations, and window result (K repetitions per arm) | `bc0219a004a0dc75` |
## All-reduce by step position

Measurement of the 91 all-reduce calls (`cross_device_reduce`) across 62 complete decode steps in the profiler trace (rank 7 of 8). Most of the wait is concentrated in the first two calls of each step (alignment of the 8 ranks), while the remaining 89 positions sit near the latency of a small intra-node all-reduce (~5.8 µs). All of it was measured under the profiler; how much of the wait remains without it was not measured.

| Position | n | Median (µs) | Mean (µs) | Min (µs) | Max (µs) | Over 50 µs | Over 20 µs |
|---|---:|---:|---:|---:|---:|---:|---:|
| Position 1 | 62 | 1,014.4 | 1,305.9 | 193.8 | 5,896.1 | 62 | 62 |
| Position 2 | 62 | 242.3 | 245.6 | 180.4 | 329.3 | 62 | 62 |
| Positions 3–91 | 5,518 | 5.79 | 6.14 | 4.86 | 826.1 | 2 | 2 |

Note: measured under profiler (tracing with instrumentation overhead elongates the step; wait without profiler was not measured). Correlation between step period and sum of all-reduces: 0.984.