Ricardo Vieira

Include Op in message when raising NotImplementedError from grad

5

### Description When grad fails because a specific Op doesn't support it it's quite difficult to figure out which one it was without jumping on an interactive debugger https://discourse.pymc.io/t/avoiding-non-nuts-samplers-when-broadcasting-a-multiplication-operation/14245/4?u=ricardov94 We...

Simplify dots with 1

3

### Description We have a `local_0_dot_x` that removes useless dots with zero'd inputs. We don't seem to have anything for dots with ones as reported in https://github.com/pymc-devs/pytensor/discussions/637#discussioncomment-8405862 ```python import pytensor...

### Description If a user wants to operate with the non-default output of RVs, they have to do `rv.make_node()` or `rv().owner.outputs` none of which is ideal or very readable. I...

### Description RandomVariables third input is a numerical integer than indexes into a list of predefined types to indicate if a variable is float64 or float32, ... This input is...

### Description If we only have on instance of the same OpFromGraph Op we should move constant inputs inside, to allow a simpler internal graph at compilation time. We shoud...

Add type hints / documentation for `mode` argument

1

### Description As discussed in https://github.com/pymc-devs/pymc/pull/7212#discussion_r1553967791 it's not clear what the signature / type is. Besides a Mode object there is a set of strings that are allowed: https://github.com/pymc-devs/pytensor/blob/a76172ee6a9753266150fc7bac4dd66906967a26/pytensor/compile/mode.py#L484-L534

### Description Now that #306 is merged, there are a couple of follow ups we should do: 1. Blockwise more Ops (Everything in linalg?) 2. Dispatch vectorize for more Ops...

## Description Marking this as major, because functionality elsewhere (PyMC) relied on this interpretation Marking as draft as it includes first commit of #664 ## Related Issue - [x] Closes...

`params_broadcast_shapes` logic is wrong

1

This helper uses `max` to broadcast dimensions, but this is wrong because a dim of length 1 should broadcasted to length 0 if that's the length of the other dim...

### Description These means we cannot use Blockwise / vectorize trivially (in the cases that would be valid). ```python import pytensor import pytensor.tensor as pt from pytensor.graph.replace import vectorize_graph x...