ComfyUI Node
Hy-Nomas Load DiT
Loads the Hy-Nomas world-model DiT (Tencent's INT8 distilled file, or a bf16/fp32 one) from models/diffusion_models.
Hy-Nomas Load DiT
- model
◄dit_name▾►
◄matmulint8►
◄quantize_floattrue►
CategoryHy-Nomas
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| dit_name | COMBO | e.g. diffusion_pytorch_model_ema_0_995_int8.safetensors (DiT/distilled_int8) | |
| matmul | COMBO | int8 | int8: the int8 tensor cores (W8A8, activations quantized per token and 128 channels; several times faster on RTX 30/40). exact: bf16 matmuls on the dequantized weights, Tencent's arithmetic bit for bit. |
| quantize_floatopt | BOOLEAN | true | For a bf16/fp32 file: quantize its linears to Tencent's int8 format at load (9 GB instead of 17 GB). |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |