Andrew James comments

Results 34 comments of


                                            Andrew James

[dynamo] Support numpy.dtype

> I'm surprised that this doesn't fix tests in test/torch_np tho, as np.dtype is used quite a bit there. Can you check what's going on? Where np.dtype was used the...

[dynamo] Handle np.iinfo/finfo/dtype as input

@pytorchbot merge

[dynamo] Handle np.iinfo/finfo/dtype as input

@pytorchbot merge

Inconsistency between constants as arguments and captured globals

NB: Also working on fixing inductor's codegen for this situation [pytorch-126195](https://github.com/pytorch/pytorch/pull/126195)

Inconsistency between constants as arguments and captured globals

@alexbaden that patch would fix the reproducer I wrote, but that was more of a demonstration rather than an exhaustive set of tests. The issue I have is really with...

[traced-graph][sparse] propagate compressed sparsity

@aartbik Pearu is on holiday, so he might not be very responsive. I took a look at your failing tests and the mkldnn failure is popping up because: 1) FakeTensors...

`pow`: fix meta function output argument dtype check.

I agree that the behavior of running the entire op in the wider dtype does not seem to match the spec outlined in the developer FAQ, however I suppose it...

RFC: In-memory sparse array interchange

> PyTorch has sparse formats that binsparse specifications does not specify as pre-defined formats. For instance, there are BSR/BSC (blocked CSR/CSC) formats, hybrid sparse formats (COO/CSR/CSC/BSR/BSC with values being strided...

Fix constant propagation in builtins and UserClasses

@atalman is someone from torchrec going to be able to help resolve this?