The inconsistency issue in dpgen results

[li-shining]

When I run the methane example from the official website, why do the results from two consecutive runs of the same task on the same node show significant discrepancies?(dpgen.log)

[njzjz-bot]

Two clean DP-GEN runs are not expected to be bitwise reproducible.

During task generation, DP-GEN replaces the descriptor, fitting-network, type-embedding, and training seeds with new random values for every model (source). It also generates a random LAMMPS velocity seed for each new exploration input (source). Therefore, using the same node does not give the same initial weights, velocities, trajectories, model deviations, or selected FP frames. GPU execution can add further nondeterminism.

This stochastic independence is intentional for the model ensemble. Compare statistical convergence and validated properties rather than exact dpgen.log counts. If an exact rerun is required, preserve and rerun the already-generated input.json and input.lammps task files with their recorded seeds; merely putting fixed seeds in default_training_param is insufficient because DP-GEN overwrites them during generation. Backend-specific deterministic GPU settings may also be required.

Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh