Ricardo Vieira

### Description Brought up in https://github.com/pymc-devs/pytensor/pull/858 ```python import pytensor import pytensor.tensor as pt a = pt.matrix("a", shape=(200, 300)) b = pt.matrix("b", shape=(300, 400)) dot = (a[:, :, None] * b).sum(1)...

Scipy 1.14.0 changed the upcasting rule of several special functions when passing integers

1

### Description Functions like `scipy.special.psi` now always upcast to float64, even if the input is a low precision integer like `int8`. We need to handle these types, but: 1) float64...

Softmax fails with integer dtypes only at runtime

1

### Description Brought up in #846 ```python import pytensor import pytensor.tensor as pt x = pt.vector("x", dtype="int64") out = pt.special.softmax(x) # Doesn't seem right out.dprint(print_type=True) # Softmax{axis=None} [id A] #...

### Description It's rather hard to find what functions are available in PyTensor (like do we have anything for linalg?). We should get sphinx to auto-populate most things with the...

Use `expand_dims` / `squeeze` in JAX implementation of `Dimshuffle`

2

https://github.com/pymc-devs/pytensor/blob/d3bd1f15a497c05a979a8e3e8be40883f669a0b6/pytensor/link/jax/dispatch/elemwise.py#L72-L89 The JAX docs of lax.reshape (which np.reshape uses) suggest this may be better for further optimizations: https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.reshape.html#jax.lax.reshape Relevant part: > For inserting/removing dimensions of size 1, prefer using lax.squeeze...

### Description With static types it is possible to find out cases where we know the reshape is useless at compile time. We should remove it when that's the case....

### Description Or have rewrites that handle this case. Otherwise it's hard / impossible to use Sparse variables in Blockwise/RandomVariables that have other batched inputs, but would otherwise work fine...

### Description AdvancedSubtensor1 and AdvancedIncSubtensor1 are specializations of AdvancedSubtensor and AdvancedIncSubtensor that only allow one vector of integer indexing along the first dimension. They exist to provide a C-implementation of...

`dvector` hardcodes `shape=(None,)`. Would be nicer to make it a partial of `vector` with the fixed dtype to still allow providing a specific static shape if the user wants to...

Default MvNormal covariance doesn't make sense

2

https://github.com/pymc-devs/pytensor/blob/5d4b0c4b9a1e478dda48e912ee708a9e557e9343/pytensor/tensor/random/basic.py#L897-L903 It is not allowed to broadcast with mu so it should be aware of its shape. And should probably be eye(mu.shape[-1]), not ones. Or we can just remove it,...