[REQ] Support for bfloat16, fp8, fp4

[cankvp]

Description

Hi, are there plans to support bfloat16, fp8, fp4?

Context

Numerical stability from bfloat16 and performance boosts from fp8, fp4

[mmacklin]

Yes, we plan to support tf32, bf16, fp8, fp4 in roughly that order. Please add any additional information on your workload / GEMM requirements if you have them. Thanks!