0.5.58: the DiT GEMM on tensor cores, 20.3 -> 14.7 s

version fix

0.5.57: the Vulkan render is 78.9 -> 20.3 s

Measured against the diffusers reference on one machine

Side-by-side against the diffusers reference, same noise

Sample rendered by 0.5.54 on an M4

Measured speed on three machines; Vulkan needs 0.5.54

q4tp layout: same 4-bit grid, predicted scales, 5.2% smaller

q4tp layout: same 4-bit grid, predicted scales, 5.2% smaller

q4tp layout: same 4-bit grid, predicted scales, 5.2% smaller

card: q4tp is the file to take

q4tp layout: same 4-bit grid, predicted scales, 5.2% smaller

Lumina-Image 2.0 q4t: text encoder + Next-DiT + VAE + tokenizer in one 3.2 GB file

model card + sample render

initial commit