FastLDF works with most things and #1113 includes tests to make sure that it behaves the same way as LogDensityFunction, which is reassuring. But there are still some outstanding things.
| Done |
TL;DR |
Issue |
PR |
| ✅ |
Make InitContext fast too |
- |
#1125 |
| ✅ |
Construct ParamsWithStats from vector + FastLDF |
#1119 |
#1129 |
| ✅ |
Provide a way to opt out of TSVI |
#1086 |
#1128 |
| ✅ |
Figure out why FastLDF is slow with untyped VarInfo |
#1113 (comment) |
Done in #1113 (86d8a73) |
| ✅ |
Use Fast InitContext for logjoint, returned, predict, etc. |
- |
#1130 |
| ✅ |
Allocation tests |
#1126 |
Done in #1113 (4b324b0) |
| ✅ |
Correctness tests for thread-safe evaluation |
- |
#1132 (6467279) |
| ✅ |
Maybe unify FastEvalVectorContext and InitFromParams. This is doable but a bit of a faff because the former handles both linked and unlinked values, whereas the latter only handles unlinked values (we also need to figure out logjac for all four possibilities). Personally, I think this is not mandatory for v0.39 as long as we keep FastEvalVectorContext internal. |
- |
#1132 |
FastLDF works with most things and #1113 includes tests to make sure that it behaves the same way as LogDensityFunction, which is reassuring. But there are still some outstanding things.
ParamsWithStatsfrom vector + FastLDFlogjoint,returned,predict, etc.