Skip to content

Release 2026-07-14 (i4 97901d3959)#131

Open
yy-code-nv wants to merge 3 commits into
mainfrom
release/2026-07-14-97901d3959
Open

Release 2026-07-14 (i4 97901d3959)#131
yy-code-nv wants to merge 3 commits into
mainfrom
release/2026-07-14-97901d3959

Conversation

@yy-code-nv

Copy link
Copy Markdown
Collaborator

Forward release from imaginaire4 to cosmos-framework: mapped file moves + import-path rewrites only, no code-level changes.

513 mapped entries — 63 changed, 4 new.

Heads-up on DMD2 / cosmos3_edge (24 files): these arrived on main via the direct private→public sync (#122), bypassing the release pipeline. This release adopts them into the mapping so the pipeline is their source of truth — so you will see pipeline-produced versions overwriting the synced copies. Diff for those files should be import-path/header normalization only.

New files: model/tokenizer/evaluation/lpips_cache.py, model/generator/mot/inference_text_kv_memory.py, model/generator/mot/unified_mot_test.py, model/generator/reasoner/qwen3_vl_moe/cosine_router_test.py.

🤖 Generated with Claude Code

Cosmos-Framework Release Bot and others added 3 commits July 23, 2026 04:17
Forward release from imaginaire4 to cosmos-framework: mapped file moves +
import-path rewrites only, no code-level changes.

Adopts the DMD2 distillation (projects/cosmos3/interactive/*) and
cosmos3_edge files into the release mapping. These also arrived on main
via the direct private->public sync (PR #122); this release makes the
mapping pipeline their source of truth, so pipeline-produced versions
overwrite the synced copies. Also ships evaluation/lpips_cache.py.

513 mapped entries; 63 changed, 4 new.
Two integration bugs surfaced only on this release branch:

- transfer.py: control-CFG called OmniMoTModel._get_velocity() without the
  newly-required keyword-only `has_noisy_actions`. Pass False — this path is
  video-transfer only (action/sound are rejected earlier in the builder).
  Fixes generator-inference-smoke (omni + multi_control_transfer).

- convert_model_to_diffusers.py: the disabled-quantization export check did an
  exact-dict compare and rejected the newly-added, inert `fp8_granularity`
  field when method=None. Ignore that field before comparing.
  Fixes generator-training-regression (nano diffusers export).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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