eval

Today we need to create a temp buffer `v` every time before using `sgmv_shrink` like ```python v = torch.zeros((x.size(0), lora_rank), dtype=x.dtype, device=x.device) _kernels.sgmv_shrink(v, x, wa_ptr, s, tmp1, layer_idx) _kernels.sgmv_cutlass(y, v,...