How to ensure full reproducibility across Colab sessions with L4 instances?

[violasiii]

I'm trying to achieve complete reproducibility when running Meridian on Google Colab with T4/L4 GPUs.
While setting random seeds like this
mmm = model.Meridian(input_data=data, model_spec=model_spec)
np.random.seed(SEED)
mmm.sample_prior(3000,seed=SEED)
mmm.sample_posterior(n_chains=7, n_adapt=3000, n_burnin=3000, n_keep=3000, seed=SEED)
ensures consistent results within a single Colab session, I get different outcomes after disconnecting/reconnecting to a new L4 instance.

[cpulavarthi]

Hello @violasiii,

Thank you for bringing this to our notice. Our team is currently investigating this issue. You may keep an eye on the Change Log which shall be updated with all the changes to Meridian including when this issue is fixed.

Thank you

Google Meridian Support Team

[yeonjaepark]

Hi @violasiii,

Reproducibility under XLA, which is used for sample_posterior(), is only possible using a stateless seed (tuple), according to tensorflow PRNG docs. To reproduce the same results from sample_posterior(), please use a tuple seed, such as (42, 42).

We are removing the internal usage of stateful seeds in sample_posterior and will be converting stateful seeds into stateless seeds internally to allow reproducibility with an int as well. You can keep an eye on the CHANGELOG to see the updated version with this change.