15B-param single-stream transformer for daVinci-MagiHuman: 40 layers,
hidden=5120, head_dim=128, GQA num_query_groups=8. Joint audio-visual
denoising in a unified token sequence (no cross-attention).
Pure addition; the DiT is dead code on the stack tip until the activation
switch in 8/8.
Files:
- fastvideo/models/dits/magi_human.py (architecture port)
- fastvideo/configs/models/dits/magi_human.py (MagiHumanVideoConfig)
- fastvideo/configs/models/dits/__init__.py (+1 export)
- tests/local_tests/helpers/magi_human_upstream.py (loader for the
upstream daVinci-MagiHuman reference - shared by parity tests)
- tests/local_tests/magi_human/test_magi_human_parity.py (base, bit-exact)
- tests/local_tests/magi_human/test_magi_human_distill_parity.py
(DMD-2 distill, bit-exact)
Two debug-scratch files were intentionally dropped during decomposition:
_debug_magi_human_block_parity.py and _debug_magi_human_weight_diff.py.
Magi-Stack: 3/8
tests/local_tests/ holds local-only parity, smoke, and component tests
organized by model family (e.g. tests/local_tests/ltx2/,
tests/local_tests/stable_audio/). They typically require a checked-out
reference repo and/or model weights and are skipped in CI; see each family's
README.md for setup. The top-level
tests/local_tests/README.md is the family index.
The CI-backed test suite still lives in fastvideo/tests/.
SolitaryThinker
4da00b35f1