Release 2026-07-27 (i4 a64e2735)

[kane-vln]

Summary

• release the current mapped Cosmos Framework tree from i4 commit a64e2735cce69f9fd388c3e19e4e04d22db7859b
• include the portable distillation-export configuration fix for VAE and Qwen paths
• add FlexAttention support and the other mapped updates recorded in .file_mapping.json

Why

Distillation export previously serialized machine-local VAE and Qwen configuration paths. Those artifacts could fail during unmodified inference on a node without the original filesystem layout. The exporter now normalizes shipped model variants to registry/package-relative paths while preserving the selected Qwen model size.

User impact

Student-only distillation exports no longer depend on the exporting machine's /lustre layout. Existing i4 export behavior and model-variant selection are preserved.

Validation

• full T2I and I2V cookbook chain on 8 nodes x 4 GPUs:
  • fresh training through iteration 5
  • strict resume through iteration 6
  • student-only export with 27 safetensors shards per model
  • unmodified four-step inference from each completed export
• final export configs contain:
  • pretrained/tokenizers/video/wan2pt2/Wan2.2_VAE.pth
  • cosmos_framework/model/generator/reasoner/qwen3_vl/configs/Qwen3-VL-32B-Instruct.json
  • no machine-local absolute paths
• T2I inference completed and produced an inspected 1024x1024 image
• I2V inference completed and fully decoded 121 frames at 736x544, 30 fps
• latest i4 release-tool tests: 69 passed
• focused public tests: 32 passed, 4 skipped
• GPU FlexAttention tests: 14 passed
• changed-file ruff, format, pre-commit, and gitleaks checks passed

Provenance

• _source_commit: a64e2735cce69f9fd388c3e19e4e04d22db7859b
• _dest_commit: 2190a72a5799ab1bc8e4ddc19ae09cd1f14d7cbc