magpie: Qwen-Image 2.1 Turbo DiT, prepacked (NVFP4 + MXFP8 + bf16)

A quantization of the transformer in Qwen/Qwen-Image-2.1-Turbo, for magpie. License: Qwen Research License (non-commercial use only), as the source model; see NOTICE.

Built with Qwen. Modified: these are not the original files. The transformer's weights were quantized and repacked; the originals are in the source repo.

One safetensors file per build, read by magpie (magpie.server.load) on sm_120 cards:

• MLP hidden and output matmuls: NVFP4 (e2m1 values, one fp8 scale per 16)
• attention q, k, v, out: MXFP8 (e4m3 values, one power-of-two scale per 32), stored as bytes under <name>@f8
• everything else: bf16

Scales are in CUTLASS's slot order for 128-wide tiles. The VAE and text encoder still come from the base repo.