[FEAT] Add MiniMax H3 joint audio-video generation - #23

Merged

lzx1413 merged 24 commits into

mainfrom

Aug 5, 2026

Commits on Aug 4, 2026

feat(minimax-h3): add model components
lzx1413
committed
4bcd205
feat(minimax-h3): add generation pipeline
lzx1413
committed
a66780e
test(minimax-h3): add parity validation harness
lzx1413
committed
26226f5
docs(minimax-h3): add usage and acceptance guide
lzx1413
committed
aa8fd88
test(minimax-h3): fix pytest module collection
lzx1413
committed
209ca87
feat(minimax-h3): complete runnable examples
lzx1413
committed
1a79aa1
docs(minimax-h3): document complete example workflows
lzx1413
committed
11a0313
fix(minimax-h3): align probed media durations
lzx1413
committed
1018b12
refactor(minimax-h3): trim resolved request state
lzx1413
committed
22e1d48
refactor(minimax-h3): separate modality VAE stages
lzx1413
committed
b9deba1
perf(minimax-h3): reuse denoising loop buffers
lzx1413
committed
35cb291
refactor(minimax-h3): share validation hashing helpers
lzx1413
committed
db91e93
refactor(minimax-h3): remove dormant VAE options
lzx1413
committed
89f9695
feat(minimax-h3): record stage execution timings
lzx1413
committed
140fd9e
fix(minimax-h3): support direct validation scripts
lzx1413
committed
641e742
feat(minimax-h3): standardize example serve entrypoints
lzx1413
committed
610a766
docs(minimax-h3): document serve workflows
lzx1413
committed
ef6a231

Commits on Aug 5, 2026

perf(minimax-h3): optimize resident multi-gpu inference
lzx1413
committed
ed64263
perf(minimax-h3): bypass parent tensor transfers
lzx1413
committed
57c6dca
docs(minimax-h3): consolidate model documentation
lzx1413
committed
202bb57
fix(minimax-h3): protect transported token metadata
lzx1413
committed
98cd34a
test(minimax-h3): add four-gpu example regression
lzx1413
committed
807730b
fix(ops): guard indexed modulation kernels
lzx1413
committed
03b3d5a
test(minimax-h3): strengthen regression acceptance
lzx1413
committed
81e861f