1setitem slot update (9 tensors, one expert 3.4 MB into 216 MB buffers): 0.305 ms2gather_qmm 1 token top-8 (3 kernels): 0.260 ms per layer3per-expert quantized_matmul loop (24 kernels): 0.428 ms per layer4stack 8 experts + gather_qmm: 0.502 ms per layer5mx.load lazy header: 10.6 ms, 9527 tensors6lazy expert load (page-cached file): first 0.528 ms, repeat 0.006 ms7