Skip to content

v2.4.0

Choose a tag to compare

@josephdviviano josephdviviano released this 20 Mar 02:15
· 253 commits to master since this release
a5649ce

What's Changed

Here's a summary of the torchgfn changes heading into v2.4.0:

Major Features

  • Relative Trajectory Balance (#457, @josephdviviano) — new loss objective added to the library.
  • Conditional GFlowNets refactor (#431, @hyeok9855) — reworked how conditioning works end-to-end, including renaming conditioning → conditions (#430) and placing conditions within States (#445).
  • torch.compile utility & benchmark harness (#443, #460, #481, @josephdviviano) — deterministic-mode toggle (#444), debug-mode fast paths for states/actions (#449), and a benchmark suite for tracking performance.
  • Local replay buffer accumulation (#446, @chirayuharyan).
  • Optimized loss computation (#451, @josephdviviano) — restructured the GFlowNet loss path for efficiency.

Minor Features & Fixes

Contributors

  • @josephdviviano — broadest footprint: compile/benchmark infra, RTB loss, loss optimization, box env, device handling, coverage, docs.
  • @younik — logging, mask caching, CI.
  • @hyeok9855 — conditional GFlowNet redesign, DB refactor, container tests, discrete env fixes.
  • @chirayuharyan — replay buffer accumulation.
  • @ali-m07 — assertion refactor.
  • @Copilot — automated fixes (L1 distance, MLP bounds, codecov config, OUTPUT_DIR).

Full Changelog: v2.3.1...v2.4.0