Skip to content

V0302: possible overflow in number format

A tensor whose static magnitude bound exceeds the largest finite value of a non-scaled format (for example 448 for fp8_e4m3) is being quantized. Scale it first or use a block-scaled format (mxfp4, nvfp4, int8).

fn main() { let t = rand_uniform([4], -1000.0, 1000.0, 1)
let q = quant(t, fp8_e4m3) }
error[V0302]: values up to 1000 may overflow fp8_e4m3 (max 448)
--> example.veer:2:10
|
2 | let q = quant(t, fp8_e4m3) }
| ^^^^^^^^^^^^^^^^^^
= note: scale the tensor down first, or use a block-scaled format (int8, mxfp4, nvfp4)
= help: run `veer explain V0302` for details