1234device Apple M5; Q2_0/Q2_0, T 256, 64 experts, tile TM 16 TN 64 TK 64 SG 2plan: 190 tiles of 16, 3040 padded rows for 2560 entriesparity: plan ok; y vs fp16-operand reference rel L2 2.03e-06, vs exact float reference 2.37e-04 PASS
1234device Apple M5; Q2_0/Q2_0, T 256, 64 experts, tile TM 16 TN 64 TK 64 SG 2plan: 190 tiles of 16, 3040 padded rows for 2560 entriesparity: plan ok; y vs fp16-operand reference rel L2 2.03e-06, vs exact float reference 2.37e-04 PASS