Skip to content

Release 2026-10-24 (i4 b3866eb3) - #309

Open
yy-code-nv wants to merge 2 commits into
mainfrom
release/2026-10-24-b3866eb3
Open

yy-code-nv wants to merge 2 commits into
mainfrom
release/2026-10-24-b3866eb3

Conversation

@yy-code-nv

Copy link
Copy Markdown
Collaborator

Automated release from i4.

_source_commit: b3866eb30721bf68dce73380563235d39fa5a622-dirty
_dest_commit (base): bf492fe2fda2108aae9b9561dc9fb9eec3fe2027

lfengad and others added 2 commits October 24, 2026 15:54
_source_commit: b3866eb30721bf68dce73380563235d39fa5a622-dirty
_dest_commit:   bf492fe
distillation_export_test.py imported cosmos3.common.public_model_config,
which only exists in i4, so unittest failed to collect the module. The
test checks restore under the i4 projects/ layout and does not apply here.
Restore the file to main; i4 now keeps the test out of the release, so the
next release produces this file unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@lfengad

lfengad commented Oct 25, 2026

Copy link
Copy Markdown
Collaborator

Release notes: default-path behavior changes in this sync

Most of this release is new opt-in features (off by default), refactors and tests. The following changes affect existing code paths.

Behavior changes

  1. Reasoner audio: with model.config.sound_und=True, video soundtracks are no longer extracted by default. Set sound_und_config.extract_audio_from_video=True to restore it. Standalone audio inputs are unaffected.
  2. Multiview attention (joint_attn_implementation="multiview") now routes from the per-camera, LiDAR and action layout metadata carried by each packed sequence:
    • Packs without that metadata use dense attention.
    • Sound packs that carry layout metadata now raise an error instead of silently falling back to dense.
    • Packs from the shipped multiview datasets carry the metadata and are unaffected.
  3. shift_lidar applies per sample: LiDAR-only samples in mixed batches now get the configured shift_lidar. Previously only all-LiDAR batches did. This affects recipes that set shift_lidar, such as the transfer-auto recipes.
  4. RandomMixedJointDataLoader:
    • Deferred samples and the lookahead limit are now tracked per source, so the realized mix follows the configured sampling probabilities.
    • Batch composition for a given seed differs from before.
    • Batches carry a _packing_budget dict.
  5. TokenMixControlConfig.log_every_n_batches removed: packing and budget statistics moved to the new monitors. Configs that set it must drop the field.

Fixes that change results in specific cases

  • EveryNDrawSample: restores MoE router accumulators after sampling, so sampling no longer perturbs router updates. MoE generator runs with aux-loss-free balancing will see different training trajectories.
  • AR multiview inference: now passes view ids when num_view_embeddings > 0, matching training.
  • Rolling static caption limit: raised from 576 to 1024 tokens.
  • easy_io: only a missing boto3, botocore or multistorageclient is tolerated. Other import errors now surface.

There are no checkpoint format changes and no removed public APIs beyond item 5.

🤖 Generated with Claude Code

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants