ASKabalan/jax-fli-scaling

jax-fli scaling benchmarks

Strong and weak scaling runs of the jax-fli particle-mesh forward model and of its initial-condition gradient, on slab decompositions of up to 512 GPUs. Experiments 11-scaling and 12-scaling-gradient draw their figures from the two perf/perf_pm.csv files. The accuracy experiments are in ASKabalan/jax-fli-experiments, and the MAP and chain outputs in ASKabalan/jax-fli-sampling.

folder	content	read by
11-scaling/perf/	perf_pm.csv with the minimum wall-time and peak per-device memory of every forward run (fli-simulate --perf), plus one jax-hpc-profiler .md report and one .args.log per run	experiment 11
11-scaling/density/	the HEALPix density lightcone of each forward run, one shell_NNNN.parquet per shell	—
11-scaling/density_spectra/	the angular power spectra of those shells, one file per run (config 11-scaling-spectra)	—
12-gradient-scaling/perf/	the same records for the gradient runs: five reverse-mode adjoints and the ckpt-20 boundary check	experiment 12
12-gradient-scaling/density/	the IC-shaped gradient field written by each gradient run (fli-simulate --grad)	—

The run name carries the configuration: exp11_{strong_M<mesh>|weak}_g<GPUs>_{f32|f64} for the forward model and exp12_{strong_M1024|weak}_g<GPUs>_{rev|ckpt<N>} for the gradient. Every parquet is a serialised jax_fli.io.Catalog.

import pandas as pd
from huggingface_hub import hf_hub_download

REPO = "ASKabalan/jax-fli-scaling"
csv = hf_hub_download(REPO, "11-scaling/perf/perf_pm.csv", repo_type="dataset")
perf = pd.read_csv(csv, header=None)  # jax-hpc-profiler rows, no header line