[feat] add MXFP8 support on H3 - #1796

Merged

SolitaryThinker merged 14 commits into

mainfrom

fastvideo-h3-quant

Sep 6, 2026

Commits on Sep 5, 2026

Refactor: simplify FastH3 example inference script.

Davids048
authored and
SolitaryThinker
committed
31cc492
rename launch script.

Davids048
authored and
SolitaryThinker
committed
9dffaa0
Update profile script.

Davids048
authored and
SolitaryThinker
committed
8386845
Add profile scripts.

Davids048
authored and
SolitaryThinker
committed
540b191
add nvtx around H3 transformer blocks

Davids048
authored and
SolitaryThinker
committed
6fac641
Add MXFP8 H3 FFN inference

Davids048
authored and
SolitaryThinker
committed
f6db9c0
Remove MiniMax H3 profile scripts

Davids048
authored and
SolitaryThinker
committed
1059362
Restore FastH3 example scripts

Davids048
authored and
SolitaryThinker
committed
dd2aed7
Add support for NVFP4 on FFN in H3.

Davids048
authored and
SolitaryThinker
committed
0fe2db3
Fix: add quack kernel dependency.
Davids048
committed
511158f
Reject FSDP Inference w/ precomputed quantized weight.
Davids048
committed
a1afd62
Reject unsafe LoRA operations after weight quantization
Davids048
committed
6e6e930
Add MXFP8 parity tests.
Davids048
committed
86618cc
pre-commit fix.
Davids048
committed
8a829b1