Checkpoint stores float8_e4m3fn (standard NVFP4), not UE8M0. The shift-by-23 was misinterpreting E4M3 bytes as E8M0 exponents.
Checkpoint stores float8_e4m3fn (standard NVFP4), not UE8M0. The shift-by-23 was misinterpreting E4M3 bytes as E8M0 exponents.