vllm_omni.platforms.npu.quant.kv_quant_npu ¶ Seeded rotation matrices for MindIE-SD quantized attention. get_quant_attention_rotation cached ¶ get_quant_attention_rotation( device: device, dtype: dtype, head_dim: int ) -> Tensor Preserve Omni's fixed Walsh-Hadamard rotation without changing global RNG.