viable/strict/1789509051: [MPS] fix rand and uniform_ upper-bound rounding (#197081)

发生了什么

Fixes #197079 Mimic CPU, warp samples rounded to the excluded upper bound back to the lower bound. Reason for putting the casts on host side and not on device side is because of Metal compiler bug. Repro: using namespace metal; kernel void repro(device bfloat* out, constant float2& bounds, constant float* samples, uint i [[thread_position_in_grid]]) { bfloat value = bfloat(bounds.x + (bounds.y - bounds.x) * samples[…

摘要按规则整理自下方来源原文

来源