Yi30/hunyuan-image3-dit-attn-capture

HunyuanImage-3.0 DiT attention capture — handoff Real attention inputs (q/k/v + mask) captured from the HunyuanImage-3.0 DiT, for kernel work: choosing/implementing an attention backend that can serve this model. The mask here is structural, not a padding mask, so it is the part that constrains what a kernel can accept. What's in this bundle File Size Contents out/call0_rank0.npz 16.5 MiB prefill — q/k/v + mask out/call32_rank0.npz 14.0 MiB denoise —… See the full description on the dataset page: https://huggingface.co/datasets/Yi30/hunyuan-image3-dit-attn-capture.