NVIDIA T4G, 7.5, 15360 MiB
torch: 2.12.0+cu132
arch_list: ['sm_75', 'sm_80', 'sm_90', 'sm_100', 'sm_110', 'sm_120']
capability: (7, 5)
compute_dtype: float16
fp16 matmul ok: True

✅ torch reached the GPU and a real fp16 matmul executed.