jnp.mean(x, dtype=bfloat16) is not respected
jnp.mean(x, dtype=bfloat16) is not respected: a task in MiMo-V2.6-RL-harbor-code: MiMo-V2.6-RL Code (Harbor) (Harbor dataset). Regarding accumulation dtype for np.mean, the numpy docs say: Note that for floating-point input, the mean is computed using the same precision the input has. Depending on…
The task
Regarding accumulation dtype for np.mean, the numpy docs say: ``` Note that for floating-point input, the mean is computed using the same precision the input has. Depending on the input data, this can cause the results to be inaccurate, especially for float32 (see example below). Specifying a higher-precision…
Part of FineEnvs/MiMo-V2.6-RL-harbor-code.