Datlehehe/gradecomp

Gradient Atoms

Unsupervised discovery and steering of model behaviors via sparse decomposition of training gradients.-

Quick Start

uv sync

# Step 1: Extract per-document gradients (8 GPUs, ~170s)
torchrun --nproc_per_node=8 experiments/extract_gradients.py --n_docs 5000

# Step 2-4: Project, decompose, and characterise atoms (~25 min CPU)
python experiments/learn_atoms.py

# Step 5: Evaluate steering
python experiments/steer_behavioral_atoms.py

Requires Phase 1 outputs (LoRA adapter + EKFAC factors) — see experiments/upstream/.

Scalability evaluation (UF HiPerGator)

This fork adds a reproducible evaluation of how the pipeline scales with the number of analysed documents N, split into a one-off GPU artifact job and a re-runnable CPU sweep. WALKTHROUGH.md documents every file; this section is the operational quick start.

One-time setup

# 1. Clone wherever you like; the WORKSPACE (artifacts, caches, results)
#    lives on /blue — never in the 40 GB home quota.
export SCALEVAL_WORKDIR=/blue/thai/$USER/gradient_atoms_scaleval   # optional:
#    this is also the auto-default when /blue/thai exists (env_common.sh).

# 2. Python environment (Phase 1 and 2 share it):
module load python/3.11        # or your preferred conda setup
python -m venv $SCALEVAL_WORKDIR/venv
source $SCALEVAL_WORKDIR/venv/bin/activate
pip install --upgrade pip
# torch first, explicitly from the CUDA 12.8 index — B200 (Blackwell) needs
# cu128 kernels; a wrong-CUDA wheel fails at runtime, not install time:
pip install "torch>=2.9.1" --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements-scaleval.txt

# 3. Gated model access: accept the license at
#    https://huggingface.co/google/gemma-3-4b-it  then
cp .env.example .env           # put HF_TOKEN=hf_... inside (git-ignored)

# 4. Slurm logs directory (sbatch --output needs it to exist):
mkdir -p logs

Phase 1 — GPU artifacts (run once, ~4–8 h)

sbatch experiments/slurm/artifacts_gpu.sbatch                    # B200 (default)
sbatch --partition=hpg-turin experiments/slurm/artifacts_gpu.sbatch   # L4 instead

Produces under $SCALEVAL_WORKDIR/artifacts/: the LoRA adapter (10 MB), EKFAC eigenfactors (2–3 GB), the 5,000 x 2,228,224 float32 gradient matrix (gradients_all.pt, ~44 GB), the frozen document snapshot, an end-to-end extraction-timing CSV, and manifest.json. Every step skips itself if its artifact already exists, so the job is safe to re-submit.

Phase 2 — CPU scaling sweep (re-runnable, ~4–8 h)

sbatch experiments/slurm/sweep_cpu.sbatch

Validates the artifacts (LoRA ordering, tensor shapes, EKFAC compatibility, projection round-trip), then runs N ∈ {500, 1000, 2000, 5000} × seeds {0,1,2} — each cell a fresh process — and renders the figures. Re-running resumes: finished cells are skipped. Outputs under $SCALEVAL_WORKDIR/results/scaling/:

• stage_runs.csv — one row per stage × N × seed: wall time, peak RSS
• quality.csv — one row per N × seed: coherence stats, live/dead atoms, counts above the paper's 0.1/0.5 thresholds
• plots/runtime_vs_n.png, memory_vs_n.png, quality_vs_n.png, extraction_vs_n.png
• cells/N…_seed…/ — per-cell atom characterisations (+ atoms.pt)

Override the grid without editing code:

sbatch --export=ALL,SWEEP_N="500 1000",SWEEP_SEEDS="0" experiments/slurm/sweep_cpu.sbatch

Local smoke test (any machine): bash experiments/run.sh sweep after pointing SCALEVAL_WORKDIR at a directory with artifacts; synthetic unit tests: python -m pytest tests/.

Resource envelope

Job	Partition	GPU	CPUs	RAM	Wall (requested)
artifacts_gpu	hpg-b200 (or hpg-turin)	1	8	160 GB	24 h
sweep_cpu	default CPU	—	30	190 GB	24 h

Both jobs use --account=thai --qos=thai and stay within 1 GPU / 30 cores / 234 GB. Blue-storage footprint ≈ 60–80 GB (gradients + factors + HF cache).

Citation

@inproceedings{gradient-atoms-2026,
  title     = {Gradient Atoms: Unsupervised Discovery of Model Behaviors
               via Sparse Decomposition of Training Gradients},
  year      = {2026},
  author    = {J Rosser}
}

License

Apache 2.0