Lukas

Ok !

yes I agree that type of coercion would be perfect!

our use-case is converting a large dataset of O(10k) files O(100TB) from ROOT to parquet. It would be great to really ensurre that all files have the same low-level type...

Thanks @jpivarski - so it seems like for our usecase having a `ak.coerce_to_type` would be suffficient then and we won't need a `ak.coerce_to_form`. Perhaps @nikoladze can confirm. Or are you...

I think I'm more concerned about this setting ``` In [18]: a = ak.Array([{'jets': [{'eta': 1.0, 'phi': 2.0}], 'electrons': [{'e': 1.23, 'pid': 11}]}]) In [19]: b = ak.Array([{'jets': [{'phi': 1.0,...

right, but say I have 1000 events of type `{"jets": var * {"eta": float64, "phi": float64}, "electrons": var * {"e": float64, "pid": int64}}` some of those events could be like...

thanks @jchodera! @aclyde11 we're happy to use our Summit allocations for any workload that might be useful. We also have a lot of (non-GPU) capacity on the distributed LHC computing...

thanks @mblondel - is this something that could be in scope for `jaxpot` later on to add closed-over variables? We can certainly provide metadata which variables require diffing and which...

could we co-opt the `static_args`-like API for this.? Edit: I guess this is equivalent to the `argnums` kwarg .. so would that be sufficient? ``` def pipeline(param_for_grad, data): def to_minimize(latent,...

thanks @shoyer - can we use `clcosure_convert` now to achieve the desired behavior?