feat(minimax-h3): port the MiniMax-H3 omni-modal video+audio DiT (portable path complete; speed GPU-gated) - #26

Open

localai-bot wants to merge 79 commits into

mainfrom

feat/minimax-h3

Commits on Aug 3, 2026

feat(minimax-h3): port the MiniMax-H3 omni-modal video+audio DiT (W0-W2)
mudler
committed
3b91d96
feat(minimax-h3): bf16 production stream, request planning, GGUF arm
mudler
committed
ad6adc2
feat(minimax-h3): reimplement the audio VAE decoder (W5)
mudler
committed
55a9ed4
feat(minimax-h3): ground the NVFP4 arm and re-scope the video VAE from real manifests
mudler
committed
f46c9e8
feat(minimax-h3): port the video-VAE decoder TransformerBlock (W4, partial)
mudler
committed
06953d3
feat(minimax-h3): complete the video-VAE ViT3D decoder (W4)
mudler
committed
24fb523
feat(minimax-h3): port the encoder text tower (W3, partial)
mudler
committed
778629b
feat(minimax-h3): assemble the t2va pipeline -- the whole path composes (W6)
mudler
committed
b549a4b
feat(minimax-h3): complete the GGUF arm -- load into a runnable DiT (W9)
mudler
committed
c46b7b8
feat(minimax-h3): NVFP4 loader -- both quantized arms now load (W10)
mudler
committed
172a072
feat(minimax-h3): port the encoder vision block -- both tower cores done (W3)
mudler
committed
9d8ad9b
feat(minimax-h3): complete the encoder -- full vision tower (W3)
mudler
committed
6d167a7
feat(minimax-h3): condition-noise augmentation for fl2va/ref2va anchors
mudler
committed
c58b112
feat(minimax-h3): reference-video geometry + frame schedule
mudler
committed
0564251
feat(minimax-h3): video-VAE spatial tiling plan and seam blend
mudler
committed
798d3d8
feat(minimax-h3): presentation token tags (the fl2va vision-span override)
mudler
committed
a00bd40
feat(minimax-h3): video-VAE 3D-CNN encoder primitives (causal Conv3d, GroupNorm3D, ResnetBlock3D)
mudler
committed
eb7b6bd
feat(minimax-h3): video-VAE Downsample3D + stride in the causal Conv3d
mudler
committed
f8c578b
feat(minimax-h3): complete the video VAE -- 3D-CNN encoder level loop
mudler
committed
af7d541
feat(minimax-h3): MM processor as Qwen3-VL reuse -- all portable work is done
mudler
committed
34986b3
feat(minimax-h3): WAV audio output; pause all GPU-requiring work per directive
mudler
committed
4fd3cac
feat(minimax-h3): video output -- PPM frames + MP4 mux argv (ffmpeg stays in examples)
mudler
committed
6a301a1
feat(minimax-h3): /v1/videos request contract + async job store
mudler
committed
9c15650
feat(minimax-h3): register /v1/videos routes + the ffmpeg example — W7 serving DONE on CPU
mudler
committed
76f41c4
feat(minimax-h3): device-resident DiT forward (brick H3-2b), GPU-verified
mudler
committed
f0bd96d
feat(minimax-h3): bf16 production stream on the device forward
mudler
committed
2435b2e
perf(minimax-h3): fold the bf16 cast into the AdaLN modulate kernels
mudler
committed
8016331
perf(minimax-h3): true bf16 storage — halves activation bytes AND is ~4x more accurate
mudler
committed
8baa709
feat(minimax-h3): keep-quant GGUF arm — runs natively where FP4 tensor cores don't
mudler
committed
5458ec4
test(minimax-h3): exercise the keep-quant arm on CUDA, not just the CPU backend
mudler
committed
5cb9268
feat(minimax-h3): audio-VAE checkpoint loader — path 2 begins
mudler
committed
e2cccf7
feat(minimax-h3): video-VAE loader — and closes a step nothing was applying
mudler
committed
d0b5eaf
fix(minimax-h3): wire post_quant_conv into the t2va pipeline
mudler
committed
39399c8
feat(minimax-h3): accept a THIRD weight-norm spelling in the audio-VAE loader
mudler
committed
ef504f8
feat(minimax-h3): encoder checkpoint loader — path 2's loaders are complete
mudler
committed
5c87342
feat(minimax-h3): assembly driver + shipped VAE config parsing
mudler
committed
cfc28ca
fix(minimax-h3): run the denoise loop DEVICE-RESIDENT, not on the CPU reference
mudler
committed
25a8932
feat(minimax-h3): wire /v1/videos into examples/server
mudler
committed
53dc4ab
chore(minimax-h3): VT_H3_PROGRESS phase trace for the denoise loop
mudler
committed
70d2107

Commits on Aug 4, 2026

Commits on Aug 5, 2026