Issues

• Avoid redundant DDP gradient synchronization during accumulation with no_sync()

  m2lines/Samudra#911 · Filter by author oa-jder-bot opened on Oct 6, 2026

• Separate benchmark startup costs from realistic loader and training throughput

  m2lines/Samudra#909 · Filter by author fomo-bot opened on Oct 6, 2026

• Coarsen LLC

  m2lines/Samudra#907 · Filter by author jder opened on Oct 5, 2026
  

• Coarsened LLC

  m2lines/Samudra#902 · Filter by author jder opened on Sep 30, 2026
  

• Samudra 2 follow ups

  m2lines/Samudra#898 · Filter by author jder opened on Sep 28, 2026
  

• Loss breakdown logging drops later output states

  m2lines/Samudra#894 · Filter by author oa-jder-bot opened on Sep 25, 2026

• Start training on new cache 1x1 and 4x4

  m2lines/Samudra#887 · Filter by author jder opened on Sep 21, 2026
  

• Add weighted_std_target so variance collapse is measurable

  m2lines/Samudra#877 · Filter by author SuMayaBee opened on Sep 5, 2026

• ARM64 GPU scrub hit CUDA OOM during Trainer setup after image pull

  m2lines/Samudra#875 · Filter by author fomo-bot opened on Sep 5, 2026

• Support scientifically correct visualization on native OM4 tripolar grids

  bugenhancement
  m2lines/Samudra#869 · Filter by author alxmrs opened on Sep 3, 2026
  6

• Choose new AWS GPU instance for CI if the capacity error persists for one more week.

  m2lines/Samudra#866 · Filter by author alxmrs opened on Sep 3, 2026
  3

• Switch from post-norm layers to pre-norm layers to improve rollout skill

  m2lines/Samudra#856 · Filter by author alxmrs opened on Aug 24, 2026
  1

• Research how to create correctors for contemporary Samudra model

  m2lines/Samudra#855 · Filter by author alxmrs opened on Aug 24, 2026
  2

  

• Move relevant LLC data to Empire AI

  m2lines/Samudra#853 · Filter by author alxmrs opened on Aug 24, 2026
  4

  

• Prepare community colab for making Samudra experiments

  m2lines/Samudra#847 · Filter by author alxmrs opened on Aug 17, 2026
  1

  

• Investigating Replay Buffer to address long term rollout concerns

  m2lines/Samudra#846 · Filter by author alxmrs opened on Aug 17, 2026
  1

  

• Building tiling strategy (multi-tile blending during training)

  m2lines/Samudra#844 · Filter by author alxmrs opened on Aug 17, 2026
  3

  

• Explore Encoder / Decoder Architectures

  m2lines/Samudra#843 · Filter by author alxmrs opened on Aug 14, 2026
  

• Update default config to incorporate freshwater fluxes

  m2lines/Samudra#839 · Filter by author alxmrs opened on Aug 4, 2026

• Address outstanding issues in Metrics code, informed by viz code.

  m2lines/Samudra#837 · Filter by author alxmrs opened on Aug 4, 2026

• In metrics suite, improve handling of the common wet mask for same-grid use cases

  m2lines/Samudra#836 · Filter by author alxmrs opened on Aug 4, 2026
  

• residual-variance RMSE doesn’t seem to expose the annual uncertainty described in the PR.

  m2lines/Samudra#835 · Filter by author alxmrs opened on Aug 4, 2026
  

• Re-run training pipelines to see the effects of incorporating freshwater fluxes in our forcings

  m2lines/Samudra#833 · Filter by author alxmrs opened on Aug 3, 2026
  

• Incorporate freshwater fluxes into our forcings

  m2lines/Samudra#832 · Filter by author alxmrs opened on Aug 3, 2026

• Training on Iris

  m2lines/Samudra#831 · Filter by author alxmrs opened on Aug 3, 2026