feat: update Minimax-H3 support - #77

Merged

Komorebi623 merged 27 commits into

THU-MIG:mainfrom

Komorebi623:main

Aug 14, 2026

Commits on Aug 7, 2026

Align MiniMax H3 conditioning with sd.cpp
Komorebi623
committed
5472c67
Optimize MiniMax H3 CUDA inference
Komorebi623
committed
a761c7c
Document H3 Diffusers gap and FC1 experiments
Komorebi623
committed
19a0fc6
Document H3 RoPE tail-view experiment
Komorebi623
committed
a6296ab
Document H3 full-task FC2 experiment
Komorebi623
committed
0d8f711
Optimize MiniMax H3 audio VAE decode
Komorebi623
committed
7dec884
Add opt-in CUDA Graph support
Komorebi623
committed
9abcff5
Optimize MiniMax H3 video postprocess
Komorebi623
committed
783af57
Parallelize MiniMax H3 video conversion
Komorebi623
committed
c8e2e70
Optimize MiniMax H3 temporal VAE assembly
Komorebi623
committed
230b302
Document MiniMax H3 VAE optimization audit
Komorebi623
committed
b48a229
Document H3 decoder QKV layout constraint
Komorebi623
committed
2f74ed7

Commits on Aug 9, 2026

Commits on Aug 11, 2026

Fix MiniMax H3 Ref2VA conditioning fidelity
Komorebi623
committed
8b0ee8b
Support media files as MiniMax H3 Ref2VA videos
Komorebi623
committed
ec063f2
Update ggml for validated MiniMax H3 SM90 routing
Komorebi623
committed
0f70722

Commits on Aug 12, 2026

chore: bump ggml submodule
yiming-l21
committed
0385126

Commits on Aug 13, 2026

Support MiniMax-H3 Diffusers Q8 paths and retain benchmarks
Komorebi623
committed
2d8c93d

Commits on Aug 14, 2026

Tune MiniMax H3 runtime and full-reference loading
Komorebi623
committed
ad0e04c
Optimize MiniMax H3 Ref2VA conditioning and profiling
Komorebi623
committed
f369f75
Merge upstream main and preserve MiniMax H3 support
Komorebi623
committed
22787eb
Add duration-based MiniMax H3 generation
Komorebi623
committed
2d318cf
Merge latest THU-MIG main
Komorebi623
committed
44000f5
docs: refresh MiniMax-H3 support and benchmarks
Komorebi623
committed
fd672fe