mattergen-generate has no seeding option — sampling is non-reproducible (training CLI already supports --seed)

[hoangngochuong24947-gif]

Summary

mattergen-generate exposes no seeding option, so sampling is non-reproducible even on the same hardware/checkpoint. The training CLI already supports a seed, so this is an inconsistency in the inference API.

Code

• mattergen/scripts/generate.py main() has no seed parameter.
• The number-of-atoms prior is drawn with an unseeded global RNG: np.random.choice(...) in NumAtomsCrystalDataset.from_num_atoms_distribution (mattergen/common/data/dataset.py:330-334).
• The condition loader constructs DataLoader(..., shuffle=True) with the global default RNG (mattergen/common/data/condition_factory.py:49-54).
• The reverse-diffusion sampler consumes the global torch RNG state.

By contrast, training has a --seed argument wired into torch.manual_seed/np.random.seed/random.seed (mattergen/diffusion/run.py:167-175, 89-107).

Reproduction

mattergen-generate run1 --pretrained-name=mattergen_base --batch_size=16 --num_batches 1
mattergen-generate run2 --pretrained-name=mattergen_base --batch_size=16 --num_batches 1
diff run1/generated_crystals.extxyz run2/generated_crystals.extxyz   # always differs

There is no --seed=42 that would make run1 and run2 agree.

Expected vs actual

Expected: mattergen-generate ... --seed=42 reproduces the same structures (same num-atoms sequence and initial noises) for a fixed checkpoint and hardware. Actual: every run differs, which complicates debugging, benchmarking, and regression testing.

Proposed fix

Add a seed: int | None = None argument to generate.py/CrystalGenerator, and before building the condition loader/sampler call:

torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
torch.cuda.manual_seed_all(seed)

Optionally thread the seed into NumAtomsCrystalDataset.from_num_atoms_distribution and DataLoader (via torch.Generator) so the num-atoms draw and the batch shuffle are reproducible too.