Not that I would understand it but I was eyeballing the logp-dlogp graphs to see where some essential looking differences occur and for instance I have seen stuff like this:
5.9.0
│ │ │ └─ Join [id BX] 229
│ │ │ ├─ 1 [id BY]
│ │ │ ├─ Composite{switch(i2, ((-1.8378770664093453 - (0.5 * i0)) - i1), -inf)} [id BZ] 227
│ │ │ │ ├─ ExpandDims{axis=1} [id CA] 221
│ │ │ │ │ └─ CAReduce{Composite{(i0 + sqr(i1))}, axis=1} [id CB] 213
│ │ │ │ │ └─ Transpose{axes=[1, 0]} [id CC] 205
│ │ │ │ │ └─ SolveTriangular{trans=0, unit_diagonal=False, lower=True, check_finite=True, b_ndim=2} [id CD] 197
│ │ │ │ │ ├─ Switch [id CE] 156
vs 5.9.1
│ │ │ └─ Join [id BX] 238
│ │ │ ├─ 1 [id BY]
│ │ │ ├─ Composite{switch(i2, ((-1.8378770664093453 - (0.5 * i0)) - i1), -inf)} [id BZ] 236
│ │ │ │ ├─ ExpandDims{axis=1} [id CA] 228
│ │ │ │ │ └─ CAReduce{Composite{(i0 + sqr(i1))}, axis=1} [id CB] 219
│ │ │ │ │ └─ Blockwise{SolveTriangular{trans=0, unit_diagonal=False, lower=True, check_finite=True, b_ndim=1}, (m,m),(m)->(m)} [id CC] 207
│ │ │ │ │ ├─ Switch [id CD] 166