Conversation
|
I need to think a bit more about this PR. Specifically, I'm concerned about the case where someone wants to generate 10000 randomizations, and they want to be able to do it in either 1 chunk of 10000, or 10 chunks of 1000, or 50 chunks of 2000, etc. It seems like the seed control of this PR doesn't go far enough. Something that might be sufficient is letting it have shape |
|
The more I think about it, the more I think that this PR isn't solving a lot of problems. It's not helping a user of |
| - 'parameter_values' <float64[2]>: Input parameter values to use during sampling. | ||
| >) | ||
| - 'parameter_values' <float64[2]>: Input parameter values to use during sampling. | ||
| <BLANKLINE> |
There was a problem hiding this comment.
what's this about? seems like a weird artifact
There was a problem hiding this comment.
it's because the __str__ of the object literally has a blank line, and apparently doctest uses this set of characters to test against it. something claude knows but i didn't.
| major versions because historically they have never, for example, changed how | ||
| ``rng.integers()`` consumes the underlying raw bit stream to generate integers, though | ||
| in principle they could. Samplomatic will issue changelog notices if it ever changes a | ||
| distribution implementation, though will avoid doing so in general. |
2 participants
Summary
This PR deprecates the
rngargument and addsseedas a 0-d tensor input.Closes #305
Details and comments
This PR also got me to think a bit deeper about what kind of seed stability we can actually support across versions and OSes. I've documented that in the sample method. Esentially, we can guarantee stability of the raw bit stream everywhere and always. However, consumers of that stream, like an implementation of generating 64 bit integers, is package dependent. NumPy treats changes to such implementations as a breaking change, so we get safety by pinning to a major version. Likewise, if samplomatic consumes these integers in a different way, that's also a breaking change to seed stability. We need to try our best not to.
AI: a bit of Claude Opus 4.6