Summary

• Add true THD packing for DeepSeek-V4 with sample-local cu_seqlens boundaries.
• Keep CP all-gather, DSV4 window/compressed attention, compressor overlap, RoPE, and TileLang indexer paths boundary-safe.
• Propagate per-sample padded lengths through response extraction, CP loss masks, log-probs, entropy, and DeepSeek-V4 assistant-only loss masking.
• Add focused packing, CP-boundary, compressor, indexer-workspace, padding-alignment, and loss-mask tests.

Validation

• Core packing revision in the target image: 7 passed, 30 warnings.
• Current rebased head 8ee08fe8: Python compileall, git diff --check, conflict-signature checks, and explicit diff review passed. The rebase only integrated upstream OPD/teacher-log-prob slicing and prompt-filter refactors in two conflict sites.
• Packing mechanics smoke: 16 nodes / 128 GPUs, 128K THD all-gather CP (TP4 / PP16 / CP2 / EP8, MBS2 / GBS2) reached Succeeded, train/loss=0.1330631522475227, and train/grad_norm=1.013722736354215.
• Checkpoint caveat: that smoke used a legacy local DeepSeekV4-pro -> FP4Unpacked -> torch_dist chain. It did not use the explicitly requested official deepseek-ai/DeepSeek-V4-Pro-Base repository, so it validates the packing path only and is not the final official-Pro-Base acceptance run.
• Official DeepSeek-V4-Pro-Base end-to-end rerun is pending conversion/loading validation.