MaksKript/NeiroHity-ACE-Step-DiT-q8

ACE-Step 1.5 turbo DiT, ONNX int8 (MatMulNBits)

Re-exported from the official ACE-Step/Ace-Step1.5 (acestep-v15-turbo/model.safetensors) DiT decoder, split into 5 chained parts (layers 0-7, 8-11, 12-15, 16-19, 20-23), weight-only int8 quantization (block 32, symmetric).

Chain: dit_L00_q8 (hidden_states, timestep, encoder_hidden_states, context_latents -> h, temb, tproj, enc) -> dit_L08_q8 -> dit_L12_q8 -> dit_L16_q8 (h, tproj, enc -> h_out) -> dit_L20_q8 (h, temb, tproj, enc -> velocity). Sequence length must be even. timestep_r = timestep (turbo).

Velocity vs official PyTorch (bf16), 250 frames: corr 0.9989 at t=1.0 and 0.5. Used by the NeiroHity Android app.