post-process-results

name	post-process-results
description	Processes and indexes OmniHub job results with omnihub-process and omnihub-index to produce index.csv and per-job processed-data. Use when the user wants to post-process results, run omnihub-process, omnihub-index, build index.csv, or work with processed-data.

Post-process OmniHub results

Order of operations

1. omnihub-process — Run first. Writes processed-data/ inside each job directory.
2. omnihub-index — Run after process. Reads all processed-data/ and writes a single table (e.g. index.csv). Fails if no processed data exists.

Results directory layout

OmniHub jobs write under a root results dir (e.g. $WORK/results/omnihub/). Each job has its own directory, typically named by $SLURM_JOB_ID:

results-dir/
  <job_id>/
    job.sh           # Copy of the submitted script
    job.yaml         # Job config (cluster, partition, app-config, etc.)
    app.yaml         # Application config used
    job-status.yaml  # Exit code, etc. (if job completed)
    logs/            # srun stdout/stderr per rank
    tools/           # omnihub-monitor, omnistat, rocprofv3-stats, etc.
    processed-data/  # Created by omnihub-process (flat YAML/JSON)

Process discovers job dirs by rglob("job.sh") under --results-dir; each such parent dir is one execution.

omnihub-process

./omnihub-process --results-dir /path/to/results/omnihub -j 4

Argument	Description	Default
--results-dir	Root results directory (contains job subdirs)	required
-j, --jobs	Parallel workers	1
--force	Re-process dirs that already have processed-data	false
--successful-executions	Skip jobs with non-zero exit code in job-status.yaml	false

Process runs a registry of parsers per job: job config, app config, app logs, job status, omnihub-monitor, omnistat variants, pytorch-trace, rccl-info, rocprofv3-stats, report card, hash. Application-specific parsers are invoked when the app has an executable applications//parse.py and the entrypoint is known (e.g. hf-finetune, vllm-latency, vllm-throughput). Parsed outputs are written under that job’s processed-data/.

omnihub-index

./omnihub-index --results-dir /path/to/results/omnihub --output index

Produces index.csv (or index.<format>). Requires that omnihub-process has already run; otherwise exits with an error.

Argument	Description
--results-dir	Same root as for process
--output	Base name for output file (e.g. index → index.csv)
--format	csv (default)

Loading the index in Pandas

The CSV uses two header rows (source and field name). Load with:

import pandas as pd
df = pd.read_csv("index.csv", header=[0, 1], index_col=0)

Application-specific parsers

If the job’s app.yaml has an entrypoint that matches a known app (e.g. applications/vllm-throughput/bench_throughput_wrapper.py), the processor looks for applications//parse.py and, if executable, runs it with the job directory as the argument. Those parsers can write additional artifacts into the job dir or into processed-data/ for inclusion in the index.