Evaluate stochastic rounding

test_sr.py

import os
import time
import torch
import torch_xla.core.xla_model as xm

N = 16

def main():
    # os.environ["XLA_USE_BF16"] = "1"
    os.environ["NEURON_RT_STOCHASTIC_ROUNDING_EN"] = "1"
    os.environ["NEURON_CC_FLAGS"] = " --internal-hlo2tensorizer-options=--experimental-unsafe-fp8e4m3fn-as-fp8e4m3 --execute-repetition=1 "

    device = xm.xla_device()
    data = torch.arange(N).reshape(1,N).expand(32,N).to(device=device) * 2
    print(f"{data=}")
    # output = (data / float(N)).to(dtype=torch.float8_e4m3fn)
    output_fp32 = (data / float(N)).to(dtype=torch.float32) * 3.11111
    print(f"{output_fp32=}")
    output_bf16 = output_fp32.to(dtype=torch.bfloat16)
    print(f"{output_bf16=}")
    output_fp8e4m3 = output_fp32.to(dtype=torch.float8_e4m3fn)
    print(f"{output_fp8e4m3=}")

if __name__=="__main__":
    main()