Run mkldnn matmul in SPR for bf32

[akote123]

In torch 2.3.0, With Reference to 0ae952d, want to mkldnn bf32 matmul in sapphire rapids but the default path is mkl. How I can enable this in SPR?

cc: @zhuhaozhe , @jgong5

Thank you

cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal @frank-wei

[jgong5]

In torch 2.3.0, With Reference to 0ae952d, want to mkldnn bf32 matmul in sapphire rapids but the default path is mkl. How I can enable this in SPR?

After you do torch.set_float32_matmul_precision("medium"), the mkldnn path will be used for float32 matmul computed with bf16.