File size: 1,059 Bytes
fa2d87b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 | # Modifications
Eligible learned linear projections in `transformer`, `transformer_ref`, and
`text_encoder` were converted from the pinned MiniMax H3 checkpoint to
OrbitQuant W4A4 packed weights. AdaLN projections use OrbitQuant INT4 group-64.
Embeddings, convolutions, norms, biases, time/context/refiner boundaries, and
policy-protected input/output projections remain in source precision.
The validated inference runtime loads the packed text encoder and selected
denoiser with BF16 compute modules so OrbitQuant dispatches every eligible
linear to the native packed W4A4 Triton path. The text encoder enters CUDA only
for conditioning, then returns to RAM before the selected denoiser enters CUDA.
The visual and audio VAEs are exact FP32 copies from the pinned source revision;
they are not quantized and are used only in a separate post-denoise decode
process. Tokenizer, processor, schedulers, license, and upstream documentation
are copied without model-weight modification. Validated examples use 50
scheduler sigma points / 49 denoiser forwards.
|