Skip to content

XLRON v1.5.1

Choose a tag to compare

@micdoh micdoh released this 12 Jul 12:01

Patch release: fix prioritized-replay resampling crash for unvmapped NUM_ENVS=1 rollouts (#56).

  • _sample_prioritized_batch assumed a (T, E, ...) leaf layout; with NUM_ENVS=1 the rollout has no env axis, so per-sample resampling crashed at the first PPO update whenever PRIO_ALPHA>0 with PRIO_BETA0 != 1.0. The trajectory (RNN) branch had the same latent assumption and wrongly-shaped importance weights.
  • NUM_ENVS>1 behavior is byte-for-byte unchanged; NUM_ENVS=1 + PER now trains.
  • New regression tests track per-sample identity through resampling for the no-env-axis layout.

Full changelog: v1.5.0...v1.5.1