gerkone/tmp-neugk-test

neugk smoke sample

Two gyrokinetic trajectories, 20 saturated frames each, in the preprocessed layout that neugk_jax (branch feat/kinetic-train) reads. It is a sample for testing the pipeline, not a training set.

part	trajectory	species	frames
kinetic/	rv1_003	ion + kinetic electron, collisional (gyaradax)	last 20 of 215, renumbered 0..19
adiabatic/	extra_iteration_0	ion, adiabatic electrons	last 20 of 267, renumbered 0..19

checkpoints/ae_hf_smoke_untrained/ holds an untrained production autoencoder (see below).

Each <name>_ifft_realpotens/ folder holds

• data/timestep_NNNNN.*.bin: df in real space, (re/im, species, vpar, mu, s, x, y), bfloat16;
• data/poten_NNNNN.*.bin: real phi (x, s, y); data/potenc_NNNNN.bin: complex phi (re/im, x, s, y), float32;
• data/linear.bin, data/linear_poten.bin: the linear eigenmode of the same configuration;
• metadata.pkl, metadata_light.pkl: grouped trajectory metadata (per-frame series sliced to the 20 frames, df window moments recomputed on them);
• config.yaml, resume/: the gyaradax config and the solver state of the last frame.

Every trajectory passes neugk_jax.dataset.schema.check_trajectory.

Setup

git clone -b feat/kinetic-train https://github.com/ml-jku/neural-gyrokinetics-jax neugk-jax
cd neugk-jax && pip install -e .
huggingface-cli download gerkone/tmp-neugk-test --repo-type dataset --local-dir ../hf_sample
export NEUGK_HF_SMOKE=$PWD/../hf_sample

Train

ae_hf_smoke is the production autoencoder (ae_full_cconv) on these two trajectories: one epoch, then validation on the same trajectories.

python main.py experiment=ae_hf_smoke output_path=runs/smoke

Evaluate a trained autoencoder

eval_checkpoint loads a run directory (its best.eqx and the normalization statistics it was trained with) and only validates it; the metrics are printed and written to <output_path>/<run id>/eval.json. checkpoints/ae_hf_smoke_untrained/ is the production autoencoder at initialization (800M parameters, model weights only), with the run config and the statistics of these two trajectories:

python main.py experiment=ae_hf_smoke output_path=runs/smoke_eval \
    eval_checkpoint=$NEUGK_HF_SMOKE/checkpoints/ae_hf_smoke_untrained

The same works for any trained ae_full_cconv run directory.

Add a gyaradax run

Run gyaradax (master) keeping every snapshot, convert the run directory and drop the trajectory next to the others. Any kinetic config on the grid of rv1_003 works; to continue rv1_003 from its last frame:

# 1. simulate: one step_*.npz per snapshot in rv1_003_gx/ (--backend jax unless the gyaradax cuda kernels are built)
gyaradax run $NEUGK_HF_SMOKE/kinetic/rv1_003_ifft_realpotens/resume/config.yaml \
    --resume-from $NEUGK_HF_SMOKE/kinetic/rv1_003_ifft_realpotens/resume/state.npz \
    --output-dir rv1_003_gx --save-dumps --n-steps 6000 --snapshot-dtype c64 --n-gpus-sp 1 --backend jax

# 2. convert: writes converted/preprocessed_kvikio/rv1_003_gx_ifft_realpotens
python -m neugk_jax.dataset.preprocess --mode gyaradax --gyaradax-dirs rv1_003_gx --target-dir converted

# 3. add it to the kinetic part and train on both
mv converted/preprocessed_kvikio/rv1_003_gx_ifft_realpotens $NEUGK_HF_SMOKE/kinetic/
python main.py experiment=ae_hf_smoke output_path=runs/smoke_gx \
    'dataset.parts.kinetic.training_trajectories=[rv1_003,rv1_003_gx]'

The converter writes the same layout as the files here (bf16 df, real and complex phi, metadata, config.yaml, resume/), with the window moments filled when the dataset first loads it. A run without saturation_time in its metadata is served from its first frame.