release: sync from i4 8ab710e8#54
Open
yy-code-nv wants to merge 5 commits into
Open
Conversation
yy-code-nv
force-pushed
the
release/2026-06-23-8ab710e8
branch
from
2f04b33 to
b9f1224
Compare
lfengad
approved these changes
Jun 25, 2026
Sync cosmos-framework with imaginaire4 commit 8ab710e8. Highlights: - Promote the new sequence_packing package (modalities, runtime, natten, packers, types, mrope, temporal_causal) under cosmos_framework/data/vfm/sequence_packing/, replacing the previous monolithic sequence_packing.py and the unified_3dmrope_utils helper (both removed). - Update consumers (mot/, callbacks/, data/vfm/, omni_mot_model.py) to the new sequence_packing API. - Update sft_dataset.py to import add_special_tokens via the new deep path (cosmos_framework.data.vfm.sequence_packing.modalities) since it is no longer re-exported from the package root. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…edMonitor use - callbacks/dit_image_sample.py: shipped (added to mapping include_files). Fixes ModuleNotFoundError in cosmos_framework.callbacks.dit_image_sample. - configs/base/defaults/callbacks.py: orphaned dataloader_speed=L( DetailedDataLoadingSpeedMonitor)(...) entry stripped. The i4 source IGNORE blocks now wrap both the import and both usage sites. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ameterize NPROC
- model/vfm/tokenizers/stable_diffusion_vae_8x8.py: shipped (fixes
ModuleNotFoundError on cosmos_framework.model.vfm.tokenizers.
stable_diffusion_vae_8x8 imported by configs/base/defaults/tokenizer.py).
- configs/base/experiment/sft/models/{nano,super}_model_config.py: drop
5 DiT-only fields (high_sigma_ratio, high_sigma_timesteps_{min,max},
use_high_sigma_strategy{,_action}) that target DiTRectifiedFlowTrainingConfig
(in the unshipped llm/dit subtree) but were being composed onto
RectifiedFlowTrainingConfig — OmegaConf rejected the unknown keys.
- tests/nano_training_smoke_test.py: parameterize on TEST_MAX_GPUS so the
smoke can run on 4-GPU allocations. Skip threshold, @pytest.mark.gpus,
NPROC_PER_NODE, and the post-train torchrun all use MAX_GPUS instead of
the hardcoded 8.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rization - cosmos_framework/model/vfm/mot/context_parallel_test.py: bring back under release authority. The IGNORE-stripped version drops the test_replicated_attention_io_cp_matches_single_rank_attention function and the two interactive.* imports it depends on (unreleased subtree). Other two tests + helpers ship cleanly. - tests/nano_training_smoke_test.py: revert TEST_MAX_GPUS parameterization back to hardcoded 8. CI server has 8 cards available so the 4-GPU workaround isn't needed. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The 2026-06-09 GB200 goldens drifted ~5e-3 in loss after the rectified-flow sigma-sampling refactor (the ``t = 1 - t_raw`` flip moved into the sampler via per-sample ``shifts``). The shift is intentional; recapture against the current code rather than restoring the legacy form. Captured 2026-06-25 on a 4 × NVIDIA GB200 node with seed 42 under ``--deterministic``. Loss reproduces bit-exact across all 10 iters; grad_norm is also deterministic here because compile.enabled=false in nano_model_config -- those values are pinned, flip to None if a future change re-enables compile and the all-rank reduction becomes non-det. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
yy-code-nv
force-pushed
the
release/2026-06-23-8ab710e8
branch
from
b9f1224 to
5a04d50
Compare
foreverlms
approved these changes
Jun 25, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Sync cosmos-framework with imaginaire4 commit 8ab710e8.
Highlights: