Skip to content

v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@hijkzzz hijkzzz released this 18 Oct 03:29
· 6 commits to main since this release
v0.2.0
baf42de
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed ?

MOLT v0.2.0 introduces a train–inference consistency skill to help diagnose and address numerical mismatches between training and rollout inference, alongside important weight refit and Docker build fixes.

Train–Inference Consistency Skill

  • Add MOLT's train–inference consistency skill for investigating numerical mismatches and working toward bitwise alignment between rollout and training execution by @hijkzzz in #209
  • Present the consistency skill as MOLT's own and cite the vLLM post by @hijkzzz in #211
  • Fix rollout dump materialization so --train.rollout_replay_dir works correctly by @hijkzzz in #212
  • Bring the train–inference consistency skill to parity with the vLLM skill by @hijkzzz in #213
  • Refine the skill with task-based guidance and consistent terminology by @hijkzzz in #214

Weight Refit Fixes and Validation

  • Fix refit to run vLLM's post-load weight processing after every weight broadcast by @hijkzzz in #217
  • Keep engine RPC execution on the main thread in the refit layout check by @hijkzzz in #222
  • Add teacher-forced refit layout validation by @hijkzzz in #223

Docker Build Fixes and Dependency Updates

  • Update AutoModel to commit b9377eee and TransformerEngine to v2.19.0 by @hijkzzz in #216
  • Fix TransformerEngine build conflicts by keeping PyTorch's CUDA 13 runtime headers out of the build include path by @hijkzzz in #218
  • Install nvidia-cudnn-frontend before building TransformerEngine 2.19 by @hijkzzz in #219
  • Align DeepEP with AutoModel's commit 10d4dd73 and patch by @hijkzzz in #221

Additional Fixes and Improvements

  • Fix rollout group-join retry behavior so consumed heavy tensors can be freed by @zhiqi-li in #215
  • Enable vLLM prefix caching when routing replay is active by @hijkzzz in #220
  • Prevent chat requests from overriding rollout temperature and other protected settings by @hijkzzz in #224
  • Simplify and clean up the implementation by @hijkzzz in #226
  • Bump version to 0.2.0 by @hijkzzz in #210

New Contributors

Full Changelog: v0.1.12...v0.2.0

Contributors

hijkzzz and zhiqi-li
Assets 2
Loading